A method for desensitizing multi-party interactions based on local differential privacy
By implementing differential privacy corpus synthesis and semantic surrogate generation within the knowledge source boundary, combined with a two-level random retrieval mechanism, the problems of insufficient privacy protection and missing semantic information in existing technologies are solved. Formal privacy guarantee and semantic information preservation are achieved across multiple rounds of queries, improving the utility of downstream tasks and reducing overhead.
Patent Information
- Application Number
- CN202611194154.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies cannot adaptively calibrate the protection strength according to the sensitivity of each knowledge source, cannot provide formal privacy guarantees while preserving the semantic information of evidence, lack a source-side privacy accounting mechanism for multi-round adaptive queries, and do not support differentiated calibration and cross-round privacy accounting for heterogeneous knowledge sources.
Differential privacy corpus synthesis is performed within the trusted boundary of each knowledge source to generate candidate surrogate corpus. A surrogate set is selected through a semantic surrogate generator, a fixed class is constructed, and a class-level retrieval index is performed. A two-level random retrieval mechanism is combined to respond to queries, ensuring privacy protection and preservation of semantic information.
It achieves continuous formal privacy guarantees under multi-round adaptive queries, avoids sensitive record inference and original text restoration, while maintaining the semantic information of downstream tasks, improving the utility of downstream tasks and reducing the overhead of multi-round queries.
Smart Images

Figure CN123069415A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of information security, specifically a method for de-identifying multi-party interactions based on local differential privacy. Background Technology
[0002] Retrieval Augmentation (RAG) significantly improves the response quality of large language models by introducing external knowledge. In recent years, RAG has been rapidly evolving into Agentic Retrieval Augmentation (AAG). However, existing technologies impose uniform protection on all knowledge sources, resulting in impaired utility of low-sensitivity sources, insufficient protection of high-sensitivity sources, and a lack of privacy accounting across multi-turn adaptive queries, making it impossible to suppress member inference and near-word-by-word extraction attacks. Summary of the Invention
[0003] This invention addresses the shortcomings of existing technologies, such as the inability to adaptively calibrate protection strength based on the varying sensitivity of different knowledge sources, the inability to provide formal privacy guarantees for private record membership relationships while preserving semantic information of evidence, the lack of source-side privacy accounting mechanisms for multi-round adaptive queries, and the lack of support for differentiated calibration and cross-round privacy accounting for heterogeneous knowledge sources. It proposes a multi-party interactive desensitization method based on local differential privacy. This method implements protection locally at each knowledge source, with strength adaptively calibrated according to the source's sensitivity and remaining privacy budget. It continuously provides formal privacy guarantees under multi-round adaptive queries. In large-model-driven multi-party data interaction scenarios (users, orchestrators, multiple knowledge source owners, and model service providers), it performs de-identification and semantic substitution processing on textual evidence flowing across trust boundaries, preventing the recipient from inferring the existence of sensitive records or restoring the original text while preserving semantic information supporting downstream tasks.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a method for desensitizing multi-party interactions based on local differential privacy. Within the trusted boundary of each knowledge source, differential privacy corpus is synthesized offline based on the private corpus of the knowledge source to obtain candidate surrogate corpus. After calibrating the class size of privacy content density awareness, a set of surrogates for each protected semantic unit is selected in groups through a semantic surrogate generator (SSG). The surrogate sets are then merged with the private blocks to obtain fixed classes. The public blocks are used as public singleton class indexes to calculate the class representation of each fixed class, resulting in a class-level retrieval index. When a subquery is received from the orchestrator in the online stage, the ordered plaintext evidence list of the orchestrator is returned through adaptive two-level random retrieval based on the subquery, fixed class components, class-level retrieval index, target return length, and online randomization parameters.
[0006] The aforementioned differential privacy corpus synthesis refers to: generating a corpus from a private corpus that matches semantically and domain distribution but does not contain any text entries that can be mapped to real private records, using a text generation method that satisfies differential privacy (including but not limited to synthesis methods based on differential privacy context learning); the synthesis process consumes a one-time synthesis budget ε. syn,i The synthesis results remain constant during the online retrieval phase.
[0007] The aforementioned knowledge source divides the corpus provided by the knowledge source owner into mutually exclusive private corpora. and public corpus The sensitivity of the source is characterized by Privacy Content Density (PCD): ; for query Semantic relevance is calculated using a public embedding model and cosine similarity: , where: D priv i D is the private corpus of the i-th knowledge source; pub i Let U be the public corpus of the i-th knowledge source, and let U be the corpus of the i-th knowledge source. Let U and U be mutually exclusive and their union be the total corpus of the i-th knowledge source. i ;ρ i Let |D| be the privacy content density (PCD) of the i-th knowledge source. priv i | represents the number of text blocks in the private corpus, |U i | represents the number of text blocks in the entire corpus; q represents the subquery sent by the editor; S(x,q) represents the semantic relevance score between text block x and subquery q; v x v q These are the embedding vectors generated by the public embedding model for text block x and subquery q, respectively; numerator v x ⋅v q The dot product of two vectors, with the denominator being the L2 norm (modulus) of the two vectors ||v||. x || and || v q The product of ‖, i.e., sim(·,·), is the cosine similarity.
[0008] The semantic surrogate generator uses a set-level exponential mechanism to select semantically similar, diverse, and secure surrogate sets for each protected semantic unit from the candidate corpus synthesized by differential privacy. Specifically, it generates a surrogate set for each protected semantic unit based on the candidate surrogate corpus, the private block of each protected semantic unit, the class size, and the utility weight.
[0009] The fixed class mentioned above contains exactly one private block from the private corpus and k. g One stand-in, totaling k g+1 member; Once the fixed class is constructed, its member composition remains unchanged during online retrieval and is only reconstructed when the corpus or publishing strategy changes. Each protected semantic unit corresponds to a fixed class, and the class content remains unchanged during online retrieval.
[0010] The two-level random retrieval refers to: the first level performs class-level retrieval based on class scores that do not rely on private source information, and optionally overlays a Bernoulli class inclusion layer; the second level performs a random response within each selected fixed class to determine whether the externally visible representative is a private block or a substitute.
[0011] The offline phase specifically includes:
[0012] S1. Differential Privacy Corpus Synthesis: Based on the private corpus D of the i-th knowledge source... priv i With the synthesis budget ε syn,i A candidate surrogate corpus P was generated using a differential privacy corpus synthesis method. syn,i , where: P syn,i ={p i,1 ,…,p i,ni} represents the set of candidate surrogate corpora; p i,1 ,…,p i,ni For n i There are 1 candidate surrogate entries, where the first index is the knowledge source number and the second index is the entry sequence number; each candidate surrogate entry is a synthesized text that does not correspond to any real private record.
[0013] This step consumes the synthesis budget in one go. The synthesized corpus remains fixed during the online phase. Candidate entries also undergo source-side publication security filtering: direct identifiers, linkable quasi-identifiers, and context-sensitive fragments are removed (following common PII / PHI de-identification and text anonymization practices). In this way, semantic similarity is used only to ensure utility, while publication security is ensured by a separate filtering process.
[0014] The differential privacy corpus synthesis method described herein employs, but is not limited to, the synthesis method based on differential privacy context learning described in "Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation".
[0015] S2. Calibrate the class size of privacy content density awareness to determine the size of the support set available for the second-level random response, specifically: based on the privacy content density ρ of the i-th knowledge source. i The set of protected semantic units G i Compared with the target density ρ set before deployment targetSelect the effective proportion of private blocks that satisfy the online release of the entire domain. The smallest positive integer class size k i ,
[0016] The effective proportion of private blocks in the online release of the entire domain. , where: |G i | is the number of protected semantic units; |D pub i | represents the number of text blocks in the publicly available corpus; (1+k) i )|G i | represents the total number of text blocks occupied by all fixed classes, i.e., one private block per protected semantic unit and k. i A stand-in; ρ target The target density has a value range of (0,1); To round up; max(1,·) indicates that the class size is at least 1; k i It is a positive integer.
[0017] The smallest positive integer class size .
[0018] The larger the class of privacy-preserving content density (PCD), the more surrogate cover is provided; the smaller the class, the closer the retrieval distribution is to the original evidence. Sources with high PCD acquire larger classes, while sources with low PCD retain smaller classes, thus achieving heterogeneous calibration. A unified approach across all sources is used by default. When more granular control is required, the strategy can also be set at the unit level. , where: k g This specifies the class size of the g-th protected semantic unit when fine-grained settings are applied; when no fine-grained settings are applied, all units in the entire source uniformly use k. i .
[0019] S3. Select substitutes in groups using a semantic substitute generator (SSG): based on the candidate substitute corpus P syn,i Private blocks c of each protected semantic unit g With class size k g ; via SSG from All Yuanzi Collection The set of surrogates S for each protected semantic unit is obtained by group sampling. g , where: g∈G i This represents the set G of protected semantic units that belong to the i-th knowledge source. i S i,kg For candidate stand-in corpus P syn,i All k g The candidate set consisting of meta-subsets, i.e., all possible candidate sets of substitutes, specifically includes:
[0020] 3.1 Candidate subsets Calculate the utility function ,in: For protected semantic units The private block, the similarity term aligns the surrogate set with the semantics of the private evidence; the diversity term Suppressing near-repetition; risk items Penalize undesirable candidates (such as class identifier fragments) under the source release strategy; weight. , Before construction, S is fixed as the set of candidate surrogates participating in the evaluation, S∈S i,kg ;U(c g S) is a private block c g c is the utility function of the candidate substitute set S at reference time; g For the protected semantic unit g, it represents the private block; Emb(·) represents the public embedding model; sim(Emb(c g Emb(s) is the cosine similarity between the embedding vectors of the private block and the surrogate entry s; (1 / k) g )∑ s∈S This indicates that for all k in set S g The average of the substitutes is calculated; Div(S) is the diversity term, which measures the degree of difference among members in set S, and can be the negative of the mean of the pairwise cosine similarity of members; Risk(S) is the risk term, which measures the severity of content in set S that violates the source publishing policy (such as class identifier fragments); λ and μ are the diversity weight and risk weight, respectively, which are non-negative real numbers and are fixed before the fixed class is constructed.
[0021] 3.2 Sample the avatar set using an exponential mechanism via SSG: ,in: For unit Budget for selecting a stand-in For utility sensitivity, Pr[G(c g )=S] indicates that the private block c g When given input, the semantic surrogate generator outputs the probability of the surrogate set S; ∝ denotes proportional to; exp(·) is the natural exponential function; ε gen,g Budget is selected for the surrogate of the protected semantic unit g; ΔU is the sensitivity of the utility function U, i.e. the maximum change in utility caused by replacing the private block; ΔU=2 is taken under the condition that the embedding vector is normalized (cosine similarity falls in [-1,1]) and the diversity term Div and the risk term Risk depend only on the candidate subset.
[0022] In this step, since the output of the semantic surrogate generator is the entire subset, each fixed class consumes the selection budget only once, rather than each surrogate consuming it once.
[0023] S4, Fixed Class Materialization: Based on the set of substitutes S for each protected semantic unit. g With private block c g Merge into fixed class E g via private source pointer θ g Record the position number z of the private block within the class. g,θg =c g And release the entire domain Ω online i Where: the fixed class E of the protected semantic unit g g =S g ∪{c g}, total k g +1 member, denoted as z g,0 ,…,z g,kg The subscript g is the unit number, the second subscript is the member number within the class, and the private source pointer θ g ∈{0,…,k g}, Release the entire domain online , ∪ and It is the union of sets.
[0024] The fixed class ;
[0025] The private source pointer is stored only within the source.
[0026] S5. Class-level retrieval index: Using public blocks as public singleton class indexes, calculate the index for each fixed class E. g The class represents μ g This means performing symmetric pooling on the embedding vectors of all members within the class.
[0027] The class mentioned above represents μ g It is fixed before online retrieval and does not depend on a private source pointer θ. g .
[0028] The online phase specifically includes:
[0029] T1. First stage: Class-level retrieval: Calculate the public singleton class score based on the subquery q and the class-level retrieval index. Score by category and fixed categories Then, the candidate class set is obtained by sorting the class scores, where: For public blocks in a public corpus, μ is the query text, i.e., the subquery, that the orchestrator sends to a knowledge source in a single interaction. g Let Emb(x) and Emb(q) be the embedding vectors of the public block and subquery, respectively; the class scores for both classes do not depend on the private source pointer θ. g .
[0030] When a class is involved in a release signal, an adaptive random inclusion layer can be overlaid: the source side only uses the public corpus to calculate common anchor points. And divide the fixed class into target classes With noise Then, Bernoulli sampling is used to determine inclusion: This layer is optional; if not used, it represents a deterministic sort. Here, τ(q) is the common anchor point, i.e., the γth percentile of the cosine similarity sim(Emb(x),Emb(q)) between all public blocks x and subqueries q in the public corpus; γ is a fixed quantile parameter before deployment, 0 < γ < 1. γ (·) indicates taking the γth percentile; G target (q) represents the set of target classes, determined by class scores h. g (q) Fixed class composition not lower than the common anchor point; G noise (q) is the noise class set, which consists of fixed classes whose class scores are lower than the common anchor point; Pr[g∈H(q)] is the probability that the fixed class g is included in the candidate class set H(q); p and r are the inclusion probabilities of the target class and the noise class, respectively, satisfying 1>p>r>0.
[0031] T2, Second Stage Intra-Class Random Response Retrieval: Based on the candidate class set and the representative budget ε of each fixed class. rep,g With class size k g Sort candidate classes by class score and retain up to a maximum of There are 10 candidate classes, of which: public singleton classes directly release their public blocks; each selected fixed class... To represent the budget Within the class, a visible representative y is sampled, specifically as follows: Private blocks may still be released—preserving the utility of precise evidence when selected by a random mechanism—but their release is no longer deterministic. Decreasing the representation budget or increasing the class size will shift the probabilistic quality toward the substitute; increasing the budget makes private evidence more likely to appear in the return list, where: K is the target return length, i.e., the maximum number of items in the return list; Pr[y=z g,θg ]=exp(ε rep,g ) / (exp(ε rep,g )+k g ) represents the probability that a private block sample is a visible representation; Pr[y=z g,j ]=1 / (exp(ε rep,g )+k g (j≠θ) g ) represents the probability that the j-th substitute sample is a visible representative; j is the member index within the class.
[0032] The probability of releasing the private block is: , where: πg priv For fixed class E g The probability of releasing a private block; ε rep,g To represent the budget; k g The size is the class size; the release probability of each of the remaining Stands is 1 / (exp(ε)). rep,g )+k g ).
[0033] T3. Post-process the visible representatives of each candidate class to obtain an ordered list of plaintext evidence L: enforce the rule of at most one representative per class, restricting repeated observations of the same fixed class; public singleton classes contribute their public blocks. The final list is sorted by class score, and if insufficient... When the item is displayed, an empty symbol is used. The padding symbol ⊥ represents an empty string and is not part of the corpus record. It is only used to fill the returned list to K entries and is ignored during downstream generation.
[0034] The orchestrator only sees the plaintext list; fixed class structures, private source pointers, and local randomness remain within the source. The source-side privacy accounting module records each release, and automatically downgrades to public-only retrieval after the budget is exhausted.
[0035] Technical effect
[0036] Compared with the prior art, the technical effects of the present invention include:
[0037] 1) Source-end fixed-class controlled release structure: Separates the release support construction from the online release decision-making, and constructs "1 private block and k" offline for each protected semantic unit. g A fixed class for releasing a "safe substitute" is used. Online, only the visible representatives within the fixed class are randomized, forming a two-layer protection structure of desensitization substitution and random release.
[0038] 2) Class size calibration for privacy content density (PCD) perception: The fixed class size is automatically derived from the proportion of protected content in the source, so that high-sensitivity sources can obtain stronger cover and low-sensitivity sources can avoid unnecessary utility loss, thus achieving heterogeneous calibration based on source sensitivity.
[0039] 3) Semantic surrogate generation (SSG) based on set-level index mechanism: The entire surrogate subset is sampled once using an index mechanism with the combined utility of "similarity + diversity - risk", and each fixed class consumes only one selection budget.
[0040] 4) Adaptive Two-Level Random Retrieval (AT-RR): The first level uses a class representation of the score that is independent of the private source pointer, and optionally overlays a Bernoulli class inclusion layer based on a common anchor; the second level performs a random response within a fixed class and restricts duplicate observations by class deduplication.
[0041] 5) Source-side session-level privacy accounting and security degradation: Offline one-time costs and online per-use costs are separated in accounting, supporting budget checks across multiple rounds of adaptive queries, and automatically switching to public-only retrieval when the budget is exhausted and no longer consuming private budget. Attached Figure Description
[0042] Figure 1 This is a flowchart of the present invention;
[0043] Figure 2 This is a flowchart of the online adaptive two-level random retrieval (AT-RR) process of the present invention;
[0044] Figure 3 This is a schematic diagram comparing the effects of embodiments of the present invention and the prior art, wherein (a) is a comparison of AUC for member inference attacks under six backbone large models (the closer to 0.5, the stronger the protection), and (b) is a comparison of BERTScore F1 for data extraction attacks under six backbone large models (the lower the score, the less leakage). Detailed Implementation
[0045] like Figure 1 As shown, this embodiment relates to a local differential privacy-based multi-party interaction de-identification system deployed within the trusted boundary of a knowledge source, including: a calibration module, a candidate surrogate corpus generation module, a fixed class construction module, a class-level indexing module, a controlled release module, and a privacy accounting module. Specifically: the calibration module calculates class size and online budget splitting based on the privacy configuration (privacy content density, target density, and session budget cap) information of the knowledge source to obtain the parameter results for adaptive two-level random retrieval; the candidate surrogate corpus generation module performs differential privacy corpus synthesis and secure filtering for publication based on the private corpus information of the knowledge source to obtain candidate surrogate corpus results; the fixed class construction module, based on the candidate surrogate... The corpus and the private block information of each protected semantic unit are sampled using a set-level index mechanism and materialized into fixed classes to obtain a fixed class set, a private source pointer, and an online release of the entire domain result. The class-level index module performs class representation calculation and index construction based on the fixed class set and public corpus information to obtain the class-level retrieval index result. The controlled release module performs two-level random retrieval and post-processing based on the subqueries of the orchestrator, the class-level retrieval index, and the parameters of the adaptive two-level random retrieval to obtain an ordered plaintext evidence list result. The privacy accounting module performs cumulative accounting and budget check processing based on the budget consumption information of each query to obtain the updated budget status result and triggers a public-only rollback when the budget is exhausted.
[0046] The specific parameters for determining AT-RR include:
[0047] 1) Cost separation. Each source records three types of costs: composite cost. Selecting the total cost with SSG One-time cost for building or refreshing a fixed class; online cost Only when using fixed-class answers to queries Billing is based on time. This separation avoids repeated invoicing of offline build costs across multiple rounds of queries, where: ε syn,i ε is the one-time synthesis budget for the i-th knowledge source; gen,g Budget selection for a substitute for the protected semantic unit g; ε gen,i =∑ g∈Gi ε gen,g The total budget for selecting a surrogate for the i-th knowledge source is the sum of the selection budgets for all protected semantic units; ε online,i,t The online budget consumed to respond to the t-th query for the i-th knowledge source, where the subscript t is the query round number.
[0048] 2) Online budget splitting. The online budget for a single query is split into... : Reserved for optional class participation layers (set to zero if not enabled), a symmetric parameterization. , Assigning to a class represents releasing it, and by default, it is at most... Each selection is evenly distributed across a fixed class, i.e. , where: ε online,i ε is the online budget for a single query of the i-th knowledge source; 1,i To reserve budget for optional class participation layers, the value is set to zero when class participation layers are not enabled; p i =exp(ε 1,i ) / (1+exp(ε 1,i )), r i =1 / (1+exp(ε 1,i )) are symmetric parameterization values for the inclusion probabilities of the target class and the noise class, respectively; ε 2,i The budget allocated to the class representative for release; K is the target return length; ε rep,g =ε 2,i / K indicates that the released budget will be evenly distributed across at most K fixed classes.
[0049] 3) Source-side accounting and rollback. Each source maintains a session limit. Answer the query Pre-inspection: If the check passes, AT-RR is executed and accounting is completed; if the check fails, a public rollback is triggered: the retrieval is limited to... No fixed class members are released, and the private source online budget is no longer consumed. Multiple knowledge sources maintain independent ledgers, and the orchestrator only receives a list of items that each source chooses to release, where: Let be the upper limit of the session budget for the i-th knowledge source; ∑ is the summation symbol, τ is the round number of the summation, and t is the current query round; D pub i For the public corpus of the i-th knowledge source, the retrieval is limited to D only when the public corpus is rolled back. pub i .
[0050] For example, a medical knowledge source corpus contains 1000 text blocks, including 800 public blocks, 200 private blocks, and 200 protected semantic units (each unit corresponds to one private block). The target density ρ is set during deployment. target =0.1, Session Budget Limit =5.0, target return length K=5.
[0051] This embodiment relates to a multi-party interactive de-identification method based on the above system, including:
[0052] In the offline phase, S1 uses private corpora as input and employs a synthesis method based on differential privacy context learning to generate candidate surrogate corpora (consuming the synthesis budget ε in one go). syn,i And filter out entries containing direct identifiers;
[0053] S2 calculates class size That is, each protected semantic unit is equipped with 5 surrogates. At this time, the effective proportion of private blocks in the online release of the entire domain is 200 / (800+6×200)=0.1, which satisfies the requirement of not exceeding the target density;
[0054] For each protected semantic unit, the semantic substitute generator samples five semantically similar but distinct substitutes from the candidate substitute corpus.
[0055] S4 merges five substitutes and one private block into a fixed class with six members and records the position of the private block using a private source pointer. Specifically, it uses a source internal vector. Record, Θ i This is a private source pointer vector within the source; This represents two proximity mappings, meaning they differ in value exactly within the same class; two mappings Proximity is defined if and only if they differ in value exactly on the same class.
[0056] S5 calculates the class representation of each fixed class and builds a class-level retrieval index.
[0057] During the online phase: Based on the subquery "medication recommendations for type 2 diabetes" issued by the orchestrator, an adaptive two-level random retrieval query is performed online with an online budget ε. online,i =1.0, Class Participation Layer (ε) is not enabled. 1,i =0), representing the release of budget ε 2,i=1.0, each selected fixed category represents a budget ε rep,g =1.0 / 5=0.2, specifically including:
[0058] T1 calculates the scores for each category and selects the top 5 categories;
[0059] T2 performs a random response for each selected fixed class: the probability of releasing the private block is exp(0.2) / (exp(0.2)+5)≈0.196, and the probability of releasing each substitute is 1 / (exp(0.2)+5)≈0.161, that is, there is about an 80% probability that the one who appears to the outside world is a substitute;
[0060] T3 performs deduplication and sorts by category score. If there are fewer than 5 results, it fills the gaps with a public null symbol ⊥ and returns a list of plaintext evidence. Simultaneously, the privacy accounting module records the entries in version 1.0. The orchestrator sees 5 readable plaintext evidence entries but cannot determine which one comes from the private corpus. When the cumulative session consumption reaches the session budget limit of 5.0, subsequent queries automatically downgrade to public-only retrieval.
[0061] The adaptive two-level random retrieval described above has the following privacy cost per stage: the class score and the common anchor point in stage one are independent of each other. Therefore, the member inference is zero-cost (0-LDP); the optional Bernoulli inclusion layer participates in channel charging as an independent class. Phase two satisfies the following conditions for each selected fixed class: -LDP: For any adjacent mappings (pointers in the same class are...) Become Output The likelihood ratio is exactly The likelihood ratio of the remaining members is 1. Through class deduplication and sequential combination, a single AT-RR call satisfies the source end... -LDP: Under adaptive multi-round queries, the session-level cumulative cost satisfies (pure differential privacy, ): ε represents the privacy budget; q represents the subquery; O represents the measurable output event; e ε The upper bound for privacy loss is given; p and r are the class inclusion probabilities; ε1 = ln max(p / r, (1-r) / (1-p)) is the cost of optional class participation in the layer; a and b are two different member position indices within the same fixed class; z g,a z g,b For the members at the corresponding positions; m is the number of fixed classes selected in this query; g j For the j-th selection, choose a fixed class; T i δ represents the total number of queries to the i-th knowledge source during the session; δ=0 indicates pure differential privacy with no relaxation term.
[0062] The cost of a session across multiple knowledge sources accumulates linearly by source and round; once the budget is exhausted, only public rollbacks do not consume any private source member budgets. Any downstream processing (sorting, formatting, generation) after returning plaintext is post-processing and does not consume additional privacy budgets.
[0063] Through practical application experiments, the multi-party interaction anonymization implementation method based on local differential privacy of this invention was run in the following experimental environment: the backbone large model adopted Qwen3 (0.6B, 1.7B, 8B, 14B), Llama-3.2-3B and gpt-oss-20b (8-bit quantized inference); the embedding model adopted Qwen3-Embedding-0.6B and all-MiniLM-L6-v2; the evaluation data adopted question answering benchmarks SQuAD and NewsQA, classification benchmarks MedNLI and AG News, and privacy corpora HealthCareMagic-100K and PubMedQA; the comparison methods were Standard RAG (ordinary retrieval baseline without privacy protection), DP-RAG, Private-RAG and VoteRAG; the attack methods were data extraction attack Spill-the-Beans and membership inference attack Riddle-Me-This; the software implementation was based on Python and PyTorch framework; the experimental parameters were set as follows: in a high privacy content density environment, ρ was set as 1. i =0.5, class size k i =4, ρ is taken in low privacy content density environments. i =0.1, class size k i =2, the corpus size for each environment is 300 text blocks; the fixed budget experiment uses an online budget ε=1.0, and the budget scanning experiment uses ε∈[0.5,1.3]; all system parameters are pre-calibrated on a non-sensitive validation set and remain unchanged during testing. The experimental data obtained are shown in Tables 1 to 5:
[0064] Table 1. Comparison of semantic utility under online budget ε=1.0 (Classification metric: MedNLI / AG News; Question answering metric: SQuAD / NewsQA)
[0065] Table 2 Comparison of Member Inference Attack (MIA) Results (Each cell represents AUC / attack accuracy; the closer to 0.5, the weaker the attack)
[0066] Table 3 Comparison of data extraction attack leaks (each cell represents ROUGE-L / BERTScore F1, with lower values indicating less observable overlap with the proprietary original data).
[0067] Table 4 Comparison of time consumption for multi-turn conversation cumulative mechanism (only including core privacy / retrieval mechanism, excluding large model generation and decoding)
[0068] Table 5. Comparison of ablation experiments under high / low privacy content density (PCD) environments (↓ indicates lower is better, ↑ indicates higher is better).
[0069] Experimental data analysis: (1) In terms of utility, according to Table 1, when the online budget ε=1.0, the F1 score of this invention on SQuAD is 0.54, which is about 45.9% higher than the optimal privacy baseline DP-RAG's 0.37; the Macro-F1 score on AG News is 0.86, which is about 30.3% higher than VoteRAG's 0.66; (2) In terms of resistance to membership inference, according to Table 2, the AUC of this invention is 0.57–0.68, close to the random guessing level of 0.5, while Standard RAG is as high as 0.87–0.97 and DP-RAG is as high as 0.96; (3) In terms of resistance to data extraction, according to Table 3, the BERTS score of this invention is F1 is 0.20–0.26, which is lower than the 0.32–0.41 of each privacy baseline; (4) In terms of system overhead, according to Table 4, the cumulative mechanism time of the present invention is lower than all privacy baselines from the 3rd round, and 1.35 seconds in the 5th round, which is lower than 1.45 seconds of DP-RAG, 1.46 seconds of Private-RAG and 1.45 seconds of VoteRAG; (5) In terms of ablation, according to Table 5, removing any component from fixed class, first-level retrieval, second-level random response, class deduplication or source adaptive calibration significantly worsens the attack indicators. For example, the MIA AUC in high PCD environment increases from 0.51 to 0.73–0.93, which verifies the necessity of each component.
[0070] Under the same online privacy budget, this invention improves downstream task utility by up to 46% compared to the optimal privacy baseline, with slower and more stable decay as the budget tightens and retrieval depth increases. The F1 score for near-word-by-word data extraction attacks is limited to 20%–26%, significantly lower than all privacy baselines. The AUC for private record member inference attacks drops to 59%–68% (close to 50% for random guessing), while the common retrieval baseline reaches 64%–96%. Under ten rounds of adaptive probing, the attack signal growth of this invention is significantly slower than variants that remove any critical component, validating the necessity of fixed classes, two-level randomization, class deduplication, and source adaptive calibration. Retrieval latency is sub-millisecond; offline construction costs are amortized over multiple rounds, and the cumulative mechanism time from the third round onwards is lower than all privacy baselines; only a fixed class index needs to be maintained at the source, resulting in manageable storage overhead.
[0071] Compared with existing technologies, the technical effects of this invention include: 1) Improved downstream task utility: When the online budget ε=1.0, the SQuAD question-answering F1 score is improved from 0.37 at the optimal privacy baseline to 0.54 (approximately 45.9%). This improvement is due to the preservation of evidentiary utility by the fixed class controlled release structure and semantic surrogate generation; 2) Improved resistance to data extraction: The extraction attack BERTScore F1 score is reduced from 0.32–0.41 to 0.20–0.26, due to the group selection of semantic surrogates and intra-class random responses; 3) Improved resistance to membership inference: The membership inference attack AUC is reduced from a maximum of 0.97 to 0.57–0.68, due to the combination of class-level retrieval, intra-class random responses, and class deduplication that is independent of private source pointers; 4) Reduced multi-round overhead: The cumulative mechanism time is lower than all privacy baselines from the 3rd round onwards, due to the two-stage design of offline construction and online reuse.
[0072] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for desensitizing multi-party interactions based on local differential privacy, characterized in that, Within the trusted boundary of each knowledge source, differential privacy corpus is synthesized based on the private corpus of the knowledge source in the offline stage to obtain candidate surrogate corpus. After that, the class size of privacy content density awareness is calibrated and the surrogate set of each protected semantic unit is selected in groups through the semantic surrogate generator (SSG). Then, the surrogate set is merged with the private block to obtain a fixed class. The public block is used as the public singleton class index to calculate the class representation of each fixed class and obtain the class-level retrieval index. When a subquery is received from the orchestrator during the online phase, the orchestrator’s ordered plaintext evidence list is returned through adaptive two-level random retrieval based on the subquery, fixed class components, class-level retrieval index, target return length and online randomization parameters. The fixed class mentioned above contains exactly one private block from the private corpus and k. g One stand-in, totaling k g +1 member; Once the fixed class is constructed, its member composition remains unchanged during online retrieval and is only reconstructed when the corpus or publishing strategy changes. Each protected semantic unit corresponds to a fixed class, and the class content remains unchanged during online retrieval.
2. The method for de-identifying multi-party interactions based on local differential privacy according to claim 1, characterized in that, The semantic surrogate generator uses a set-level exponential mechanism to select semantically similar, diverse, and secure surrogate sets for each protected semantic unit from the candidate corpus synthesized by differential privacy. Specifically, it generates a surrogate set for each protected semantic unit based on the candidate surrogate corpus, the private block of each protected semantic unit, the class size, and the utility weight.
3. The method for de-identifying multi-party interactions based on local differential privacy according to claim 1, characterized in that, The two-level random retrieval refers to: the first level performs class-level retrieval based on class scores that do not rely on private source information, and optionally overlays a Bernoulli class inclusion layer; the second level performs a random response within each selected fixed class to determine whether the externally visible representative is a private block or a substitute.
4. The method for de-identifying multi-party interactions based on local differential privacy according to any one of claims 1-3, characterized in that, The offline phase specifically includes: S1. Differential Privacy Corpus Synthesis: Based on the private corpus D of the i-th knowledge source... priv i With the synthesis budget ε syn,i The candidate surrogate corpus P is generated using a differential privacy corpus synthesis method. syn,i , where: P syn,i ={p i,1 ,…,p i,ni } represents the set of candidate surrogate corpora; p i,1 ,…,p i,ni For n i There are 1 candidate surrogate entries, where the first index is the knowledge source number and the second index is the entry sequence number; each candidate surrogate entry is a synthesized text that does not correspond to any real private record. S2. Calibrate the class size of privacy content density awareness to determine the size of the support set available for the second-level random response, specifically: based on the privacy content density ρ of the i-th knowledge source. i The set of protected semantic units G i Compared with the target density ρ set before deployment target Select the effective proportion of private blocks that satisfy the online release of the entire domain. The smallest positive integer class size k i ; S3. Select substitutes in groups using a semantic substitute generator (SSG): based on the candidate substitute corpus P syn,i Private blocks c of each protected semantic unit g With class size k g ; via SSG from All Yuanzi Collection The set of surrogates S for each protected semantic unit is obtained by group sampling. g , where: g∈G i This represents the set G of protected semantic units that belong to the i-th knowledge source. i S i,kg For candidate stand-in corpus P syn,i All k g The candidate set consisting of meta-subsets, i.e., all possible candidate sets of substitutes; S4, Fixed Class Materialization: Based on the set of substitutes S for each protected semantic unit. g With private block c g Merge into fixed class E g via private source pointer θ g Record the position number z of the private block within the class. g,θg =c g And release the entire domain Ω online i Where: the fixed class E of the protected semantic unit g g =S g ∪{c g }, total k g +1 member, denoted as z g,0 ,…,z g,kg The subscript g is the unit number, the second subscript is the member number within the class, and the private source pointer θ g ∈{0,…,k g }, Release the entire domain online , ∪ and The union of sets; S5. Class-level retrieval index: Using public blocks as public singleton class indexes, calculate the index for each fixed class E. g The class represents μ g This means performing symmetric pooling on the embedding vectors of all members within the class.
5. The method for de-identifying multi-party interactions based on local differential privacy according to claim 4, characterized in that, The effective proportion of private blocks in the online release of the entire domain. , where: |G i | is the number of protected semantic units; |D pub i | represents the number of text blocks in the publicly available corpus; (1+k) i )|G i | represents the total number of text blocks occupied by all fixed classes, i.e., one private block per protected semantic unit and k. i A stand-in; ρ target The target density has a value range of (0,1); To round up; max(1,·) indicates that the class size is at least 1; k i It is a positive integer; The smallest positive integer class size ; The larger the class of privacy content density perception, the more surrogate cover is provided; the smaller the class, the closer the retrieval distribution is to the original evidence. Sources with high PCD obtain larger classes, while sources with low PCD retain smaller classes, thus achieving heterogeneous calibration. The default is to use a unified approach across all sources. When more granular control is required, the strategy can also be set at the unit level. , where: k g This specifies the class size of the g-th protected semantic unit when fine-grained settings are applied; when no fine-grained settings are applied, all units in the entire source uniformly use k. i .
6. The method for de-identifying multi-party interactions based on local differential privacy according to claim 5, characterized in that, Step S3 specifically includes: 3.1 Candidate subsets Calculate the utility function ,in: For protected semantic units The private block, the similarity term aligns the surrogate set with the semantics of the private evidence; the diversity term Suppressing near-repetition; risk items Penalize undesirable candidates (such as class identifier fragments) under the source release strategy; weight. , Before construction, S is fixed as the set of candidate surrogates participating in the evaluation, S∈S i,kg ;U(c g S) is a private block c g c is the utility function of the candidate substitute set S at reference time; g For the protected semantic unit g, it represents the private block; Emb(·) represents the public embedding model; sim(Emb(c g Emb(s) is the cosine similarity between the embedding vectors of the private block and the surrogate entry s; (1 / k) g )∑ s∈S This indicates that for all k in set S g The average of the substitutes is calculated; Div(S) is the diversity term, which measures the degree of difference among members in set S, and can be the negative of the mean of the pairwise cosine similarity of members; Risk(S) is the risk term, which measures the severity of content in set S that violates the source publishing policy; λ and μ are the diversity weight and risk weight, respectively, and are non-negative real numbers that are fixed before the fixed class is constructed; 3.2 Sample the avatar set using an exponential mechanism via SSG: ,in: For unit Budget for selecting a stand-in For utility sensitivity, Pr[G(c g )=S] indicates that the private block c g When given input, the semantic surrogate generator outputs the probability of the surrogate set S; ∝ denotes proportional to; exp(·) is the natural exponential function; ε gen,g Budget is selected for the surrogate of the protected semantic unit g; ΔU is the sensitivity of the utility function U, i.e. the maximum change in utility caused by replacing the private block; ΔU=2 is taken under the condition that the embedding vector is normalized (cosine similarity falls in [-1,1]) and the diversity term Div and the risk term Risk depend only on the candidate subset.
7. The method for de-identifying multi-party interactions based on local differential privacy according to claim 4, characterized in that, The online phase specifically includes: T1. First stage: Class-level retrieval: Calculate the public singleton class score based on the subquery q and the class-level retrieval index. Score by category and fixed categories Then, the candidate class set is obtained by sorting the class scores, where: For public blocks in a public corpus, μ is the query text, i.e., the subquery, that the orchestrator sends to a knowledge source in a single interaction. g Let Emb(x) and Emb(q) be the embedding vectors of the public block and subquery, respectively; the class scores for both classes do not depend on the private source pointer θ. g ; T2, Second Stage Intra-Class Random Response Retrieval: Based on the candidate class set and the representative budget ε of each fixed class. rep,g With class size k g Sort candidate classes by class score and retain up to a maximum of There are 10 candidate classes, of which: public singleton classes directly release their public blocks; each selected fixed class... To represent the budget Within the class, a visible representative y is sampled, specifically as follows: Private blocks may still be released—preserving the utility of precise evidence when selected by a random mechanism—but their release is no longer deterministic. Decreasing the representation budget or increasing the class size will shift the probability quality toward the substitute; increasing the budget makes private evidence more likely to appear in the return list, where: K is the target return length, i.e., the maximum number of items in the return list; Pr[y=z g,θg ]=exp(ε rep,g ) / (exp(ε rep,g )+k g ) represents the probability that a private block sample is a visible representation; Pr[y=z g,j ]=1 / (exp(ε rep,g )+k g (j≠θ) g ) represents the probability that the j-th stand-in is a visible representative; j is the member index within the class; T3. Post-process the visible representatives of each candidate class to obtain an ordered list of plaintext evidence L: enforce the rule of at most one representative per class, restricting repeated observations of the same fixed class; public singleton classes contribute their public blocks, and the final list is sorted by class score, with insufficient... When the item is displayed, an empty symbol is used. The padding symbol ⊥ represents an empty string and is not part of the corpus record. It is only used to fill the returned list to K entries and is ignored during downstream generation. The orchestrator only sees the plaintext list; the fixed class structure, private source pointers and local randomness are all left in the source. The source-side privacy accounting module records each release, and automatically degrades to public-only retrieval after the budget is exhausted.
8. The method for desensitizing multi-party interactions based on local differential privacy according to claim 7, characterized in that, The probability of releasing the private block is: , where: π g priv For fixed class E g The probability of releasing a private block; ε rep,g To represent the budget; k g The size is the class size; the release probability of each of the remaining Stands is 1 / (exp(ε)). rep,g )+k g ).
9. The method for de-identifying multi-party interactions based on local differential privacy according to claim 8, characterized in that, When a class is involved in a release signal, an adaptive random inclusion layer can be overlaid: the source side only uses the public corpus to calculate common anchor points. And divide the fixed class into target classes With noise Then, Bernoulli sampling is used to determine inclusion: This layer is optional; if not used, it represents a deterministic sort. Here, τ(q) is the common anchor point, i.e., the γth percentile of the cosine similarity sim(Emb(x),Emb(q)) between all public blocks x and subqueries q in the public corpus; γ is a fixed quantile parameter before deployment, 0 < γ < 1. γ (·) indicates taking the γth percentile; G target (q) represents the set of target classes, determined by class scores h. g (q) Fixed class composition not lower than the common anchor point; G noise (q) is the noise class set, which consists of fixed classes whose class scores are lower than the common anchor point; Pr[g∈H(q)] is the probability that the fixed class g is included in the candidate class set H(q); p and r are the inclusion probabilities of the target class and the noise class, respectively, satisfying 1>p>r>0.
