Anti-fraud detection and prevention system in RTB bidding based on game theory

Abstract

The invention discloses an anti-fraud detection and prevention system in RTB bidding based on a game theory, and relates to the field. Comprising a plurality of demand side platforms DSPs used for submitting advertisement bids to an advertisement transaction platform ADX; the market anti-fraud manager MAM is integrated in an advertisement exchange ADX and is used for executing the following operations: receiving bids submitted by each DSP; the method is based on a repeated game model and a Nash equilibrium theory. According to the method, the fraud rate is reduced, a simulation test shows that false bids and collusion behaviors can be reduced by 20-30%, and the market fairness is improved. And dynamic prevention: guiding the long-term honesty behavior of the DSP through game excitation, which is superior to static rule detection. And engineering landing: ensuring millisecond-level response through hierarchical calculation, and adapting to a high-frequency RTB scene. And excitation compatibility: approximate Nash equilibrium is realized under repeated games, and short-term fraud incentives are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of anti-fraud detection and prevention in real-time bidding (RTB) auctions based on game theory. BACKGROUND

[0002] With the rapid development of programmatic advertising, RTB has become the mainstream mode of advertising transactions. Multiple DSPs bid for advertising display positions within milliseconds, and traffic distribution is achieved through ad exchanges (ADX, Ad Exchange). However, the RTB market is vulnerable to fraudulent behavior, including:

[0003] False bidding: DSPs submit artificially high bids to disrupt the market but have no intention of fulfilling them, increasing the cost of other DSPs.

[0004] Collusion: multiple DSPs coordinate to raise bids or reserve prices, driving up market prices to profit.

[0005] Brushing: fake clicks or conversions are used to fake ad effectiveness, misleading ROI (return on investment) evaluation.

[0006] These behaviors result in loss of platform revenue, decreased trust among advertisers, and market instability.

[0007] Existing technologies attempt to address fraud issues but have limitations:

[0008] 1. Chinese patent CN112801704A: detects abnormal traffic (such as high CTR anomalies) based on statistical thresholds and machine learning, but is a post-analysis, lacks real-time prevention, and does not consider the dynamic nature of multi-DSP interactions.

[0009] 2. Chinese patent CN113837782A: uses master-slave game theory to optimize the bidding and budget allocation of a single DSP, does not involve multi-party game or fraud detection, and is difficult to address coordinated fraud in the RTB market.

[0010] 3. Chinese patent CN114358854A: applies game theory to optimize advertising resource allocation, focusing on efficiency rather than anti-fraud, and does not provide real-time detection or incentive mechanisms.

[0011] 4. US patent US10460352B2: proposes a game-based bidding optimization, focusing on short-term auction strategies, and does not involve long-term repeated games or anti-fraud design.

[0012] Furthermore, academic literature (such as the 2023 arXiv paper "Game-Theoretic Incentives for Truthful Bidding in RTB Auctions") discusses game incentives, but does not address the millisecond-level latency or engineering implementation of RTB. Existing technologies lack dynamic anti-fraud mechanisms for multi-DSP repeated game scenarios, particularly in terms of real-time performance (<100ms), incentive compatibility, and engineering implementation. This invention aims to fill this gap by using a game theory model to provide an integrated system for real-time detection and prevention. Summary of the Invention

[0013] The purpose of this invention is to address the shortcomings of existing technologies by proposing a fraud detection and prevention system for RTB bidding based on game theory.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] A fraud detection and prevention system for RTB bidding based on game theory includes:

[0016] Multiple demand-side platforms (DSPs) are used to submit ad bids to the advertising exchange platform ADX.

[0017] Marketplace Anti-Fraud Manager (MAM), integrated into the Ad Exchange (ADX), is used to perform the following operations:

[0018] A1: Receive bids submitted by each DSP. i (t);

[0019] A2: Based on the repeated game model and Nash equilibrium theory, the deviation δ of each DSP bid relative to the estimated equilibrium strategy is calculated in real time. i (t) and its behavior cause externalities E to other market participants. i (t), where the deviation δ i (t) is compared with a preset threshold T. If it exceeds the threshold, fraudulent behavior is determined.

[0020] A3: Generate a dynamic penalty tax based on the judgment result, τ i (t)=λE i (t), where λ is the learnable penalty tax rate;

[0021] A4: Based on the final payment price p i (t)=p base (t)+τ i (t) and combined with the reputation scores of each DSP r i Adjusting the allocation of ad display positions (a) i (t);

[0022] A5: Record all bidding, detection, penalty and allocation data to an immutable audit log, and periodically perform parameter updates and backtracking settlement operations.

[0023] Preferably, the Market Anti-Fraud Manager (MAM) includes a detection module, and the working logic of the detection module is as follows:

[0024] B1: Employs an iterative optimal response algorithm as an approximate Nash solver to estimate the bid deviation δ in real time. i (t);

[0025] B2: The Isolation Forest unsupervised machine learning algorithm is used to detect anomalies in the historical bidding sequences of DSPs to help identify deviation behaviors;

[0026] B3: Calculate the mutual information S between different DSP bidding sequences ij If S ij Exceeding the experience threshold will trigger a conspiracy alert;

[0027] B4: Estimating DSP through counterfactual simulation methods i Behavior of DSP j Impact of expected returns And sum to obtain the total externality

[0028] Preferably, the Market Anti-Fraud Manager (MAM) includes a prevention module, and the working logic of the prevention module is as follows:

[0029] C1: Maintain and dynamically update the reputation score of each DSP. i , Historical performance rate is calculated based on the accuracy of budget expenditures, while fraud rate is calculated based on historical delta. i (t) statistics;

[0030] C2: Credit score i With penalty tax rate λ and allocation weight a i (t) is associated with imposing higher penalty rates or lowering the allocation priority of low-reputation DSPs;

[0031] C3: Employs reinforcement learning algorithms to ensure long-term platform benefits. k To optimize the objective, the penalty tax rate λ is dynamically updated using the formula λ←λ+η·▽ λ E[R k ], where η is the learning rate, ▽ λ E[R k ] represents the gradient, which is estimated by replaying historical logs.

[0032] Preferably, the Market Anti-Fraud Manager (MAM) includes a log auditing module, and the working logic of the log auditing module is as follows:

[0033] D1: Record the request ID and the bids of each DSP at each bidding time t. i (t), calculated deviation δ i (t), penalty tax τ i (t), final payment p i (t), allocation result a i (t);

[0034] D2: At the end of the settlement period, call the anti-fraud detection service to verify the authenticity of click and conversion data after the ad campaign;

[0035] D3: Perform retrospective settlement based on the verification results to recover improper gains resulting from fraud or compensate for misjudgments;

[0036] D4: Provides differential privacy protection for all data processing operations and provides immutability and auditability for all log records.

[0037] Preferably, the Market Anti-Fraud Manager (MAM) includes a collusion detection module, and the working logic of the collusion detection module is as follows:

[0038] E1: Based on information theory methods, compute any two DSPs i DSP j Mutual information S of bidding sequences ij ;

[0039] E2: Statistical correlation of its behavior;

[0040] E3: Set dynamic threshold T collusion When S ij >T collusion Time-determination DSP i DSP j There is potential for collusion:

[0041] E4: Implement enhanced penalties for DSPs that trigger conspiracy alerts.

[0042] Preferably, in step S4, the punitive measures include increasing the penalty tax rate λ and temporarily freezing the credit score r. i One or more of the following measures may be taken: suspending their eligibility to participate in advertising space bidding.

[0043] Preferably, the anti-fraud detection and prevention system in RTB bidding adopts a layered computing architecture to meet the low latency requirements of RTB:

[0044] Deploy lightweight detection models on edge computing nodes, utilize local caching and KNN approximation algorithms to calculate approximate values, and make real-time decisions;

[0045] A high-precision Nash equilibrium simulation model is periodically run at the central computing node to update the global game model parameters and thresholds, and then distributed to the edge nodes.

[0046] Edge nodes work in collaboration with central nodes to ensure that the entire system responds to bidding requests in less than 100 milliseconds.

[0047] Preferably, the anti-fraud detection and prevention method in RTB bidding includes the following steps:

[0048] S1: Receive bids from multiple DSPs. i (t);

[0049] S2: Based on the repeated game Nash equilibrium model, calculate the deviation δ of each bid in real time. i (t) and externality E i (t);

[0050] S3: The deviation δ i (t) is compared with the dynamic threshold T, if δ i If (t) > T, then a penalty tax τ is generated. i (t), otherwise set τ i (t) = 0;

[0051] S4: Calculate the final payment p i (t), and combine DPS reputation score to determine ad allocation a i (t);

[0052] S5: Record all relevant data for this auction to the audit log;

[0053] S6: Perform the following sub-steps periodically;

[0054] S61: Estimate platform revenue R based on audit log playback. k ;

[0055] S62: Update the penalty tax rate parameter λ using a reinforcement learning algorithm;

[0056] S63: Perform retrospective settlement and adjust revenue distribution based on actual ad conversion data.

[0057] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement steps S1-S6 of a game theory-based anti-fraud detection and prevention system for RTB bidding.

[0058] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, constitutes steps S1-S6 of an anti-fraud detection and prevention system for RTB bidding based on game theory.

[0059] The beneficial effects of this invention are as follows:

[0060] Reduce fraud rate: Simulation tests show that it can reduce false bids and collusion by 20-30%, improving market fairness.

[0061] Dynamic prevention: Guiding DSP to behave honestly in the long term through game-theoretic incentives is superior to static rule detection.

[0062] Engineering Implementation: Layered computing ensures millisecond-level response and is suitable for high-frequency RTB scenarios.

[0063] Incentive compatibility: Achieving near-Nash equilibrium in repeated games reduces short-term incentives for fraud.

[0064] Compliance: Supports privacy protection and auditing, and complies with GDPR and the Personal Information Protection Act. Attached Figure Description

[0065] Figure 1 This is a diagram illustrating the overall architecture of a fraud detection and prevention system for RTB bidding based on game theory, as proposed in this invention.

[0066] Figure 2 This is a bidding sequence diagram for an anti-fraud detection and prevention system in RTB bidding based on game theory proposed in this invention;

[0067] Figure 3 This is a schematic diagram of the fraud detection process of an anti-fraud detection and prevention system for RTB bidding based on game theory proposed in this invention.

[0068] Figure 4 This invention presents a data flow diagram of the incentive and penalty mechanisms for an anti-fraud detection and prevention system in RTB bidding based on game theory.

[0069] Figure 5 The pseudocode flowchart of an anti-fraud detection and prevention system algorithm for RTB bidding based on game theory proposed in this invention is shown below.

[0070] Figure 6 This is an example game model and equilibrium diagram of an anti-fraud detection and prevention system for RTB bidding based on game theory proposed in this invention. Detailed Implementation

[0071] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0072] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0073] Explanation of terms and symbols.

[0074] N: The set of DSPs participating in the bidding, where i∈N represents the i-th DSP.

[0075] t: Bidding time / serial number.

[0076] b i (t): DSP i's bid at t.

[0077] v i (t): The true value of DSP i (proprietary).

[0078] a i (t): Allocation instruction (1 for successful bid, 0 for unsuccessful bid).

[0079] p i (t): Final payment, p i (t)=p b ase(t)+τ i (t).

[0080] E i (t): Externality of DSP i (impact on other players).

[0081] δ i (t): Nash bias (a measure of bids deviating from equilibrium).

[0082] r i DSP i's reputation score (0-1).

[0083] λ: Adjustment coefficient (penalty tax rate).

[0084] U i : The utility function (long-term ROI) of DSP i.

[0085] Example 1: Overall Mechanism Framework

[0086] 1. System Composition (see...) Figure 1 )

[0087] DSP and ADX: Multiple DSPs interact with the ADX via the OpenRTB protocol to submit bids. i (t).

[0088] Marketplace Anti-Fraud Manager (MAM): Integrated into ADX, responsible for receiving bids, detecting fraud, and calculating penalty taxes. i (t), Adjust allocation a i (t), and periodically update the parameters.

[0089] Detection module: Combines game theory models and machine learning to analyze bid sequences.

[0090] Prevention module: manages reputation score r_i and incentive mechanisms.

[0091] Log auditing module: Records transaction logs, supporting backtracking settlement and compliance auditing.

[0092] 2. Bidding / Execution Process (see...) Figure 2 )

[0093] S1: DSP submission b i (t) and optional intentions (such as budget surplus).

[0094] S2: MAM reads historical logs and calls the detection module to calculate δ. i (t) and E i (t).

[0095] S3: If δ i If (t) > threshold T, it is marked as fraud, and a penalty tax τ is calculated. i (t)=λ·E i (t).

[0096] S4: Based on p i (t)=p b ase(t)+τ i (t)(p b ase is the payment for the second-price auction) to determine the allocation of a_i(t).

[0097] S5: Record conversion events (such as click-through rate) and perform retrospective adjustments during periodic settlement.

[0098] Example 2: Mathematical Model and Detection Design

[0099] 1. Game Theory Model

[0100] RTB is considered a repeated game, and the utility function of DSP i is:

[0101] U i =∑ t [v i (t)·a i (t)-p i (t)-c 欺诈 (t)]

[0102] Where c欺诈 The cost of fraud (e.g., reputation loss). Nash equilibrium is defined as a strategy σ* where no player unilaterally deviates from the profit-generating objective. Deviation δ i (t) is:

[0103] 2. Detection methods (see...) Figure 3 )

[0104] Real-time estimation: Calculate δ using an approximate Nash solver (such as the iterative optimal response). i (t).

[0105] Machine learning assistance: Isolation Forest is used to detect anomalies in the bid sequence, combined with mutual information S ij Analysis of collusion.

[0106] Externality calculation: Δ j The impact of behavior i on the expected return of j is estimated through counterfactual simulation.

[0107] 3. Parameter Learning

[0108] MAM uses the platform objective R_k (revenue minus fraud loss) as the reward and adjusts λ using reinforcement learning:

[0109]

[0110] Where η is the learning rate, and R0 k Estimated by replaying historical logs.

[0111] Example 3: Prevention and Balance Properties

[0112] 1. Prevention mechanism: Low reputation score i Triggering higher τ i (t) or the weighting is reduced, forcing the DSP to avoid fraud.

[0113] 2. Equilibrium Analysis: In repeated games with a discount factor δ < 1, the short-term gain from fraud is less than the long-term loss of reputation, and players tend to reach a cooperative equilibrium (honest bidding).

[0114] Example 4: Reputation and Retrospection

[0115] 1. Reputation Update:

[0116] r i = α·historical fulfillment rate + β·(1-fraud rate), α+β=1

[0117] Fulfillment rate is based on budget execution, fraud rate is based on delta. i (t) Statistics.

[0118] 2. Retrospective Settlement: At the end of the cycle, MAM adjusts τ based on conversion authenticity (combined with anti-fraud detection). i (t), return or deduct improper gains.

[0119] Example 5: Collusion Detection

[0120] 1. Metric: Calculate the mutual information between DSP i and j:

[0121] If S ij >T (experience threshold), triggering a reinforced penalty (such as freezing r). i ).

[0122] 2. Countermeasure: Increase τ i (t) or suspend DSP participation in bidding.

[0123] Example 6: Engineering Implementation

[0124] 1. Low-latency computing:

[0125] Edge nodes: δ is approximated using local caching and KNN (Nearest Neighbor) algorithm. i (t).

[0126] Central node: Periodic high-precision Nash simulation, updating λ.

[0127] 2. Privacy protection: Differential privacy is applied to the log data (Gaussian noise is added).

[0128] 3. Scalability: MAM is deployed as a microservice, compatible with OpenRTB v2.5 (the ext field carries δ). i / τ i ).

[0129] 4. Audit: All data (bids, delta) i τ i The allocation is recorded in an immutable log, supporting compliance checks.

[0130] Example 7: Parameters and Alternative Solutions

[0131] Allocation rules: Generalized second valence (GSP) or VCG variants can be used, with weights combined with r. i .

[0132] Detection variant: Supports dynamic threshold T (based on historical fraud rate).

[0133] Learning algorithm: PPO (Proximal Policy Optimization) can be used to replace gradient update optimization of λ.

[0134] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fraud detection and prevention system for RTB bidding based on game theory, characterized in that, include: Multiple demand-side platforms (DSPs) are used to submit ad bids to the advertising exchange platform ADX. Marketplace Anti-Fraud Manager (MAM), integrated into the Ad Exchange (ADX), is used to perform the following operations: A1: Receive bids submitted by each DSP. i (t); A2: Based on the repeated game model and Nash equilibrium theory, the deviation δ of each DSP bid relative to the estimated equilibrium strategy is calculated in real time. i (t) and its behavior cause externalities E to other market participants. i (t), where the deviation δ i (t) is compared with a preset threshold T. If it exceeds the threshold, fraudulent behavior is determined. A3: Generate a dynamic penalty tax based on the judgment result, τ i (t)=λE i (t), where λ is the learnable penalty tax rate; A4: Based on the final payment price p i (t)=p base (t)+τ i (t) and combined with the reputation scores of each DSP r i Adjusting the allocation of ad display positions (a) i (t); A5: Record all bidding, detection, penalty and allocation data to an immutable audit log, and periodically perform parameter updates and backtracking settlement operations.

2. The anti-fraud detection and prevention system for RTB bidding based on game theory according to claim 1, characterized in that, The Market Anti-Fraud Manager (MAM) includes a detection module, and the working logic of the detection module is as follows: B1: Employs an iterative optimal response algorithm as an approximate Nash solver to estimate the bid deviation δ in real time. i (t); B2: The Isolation Forest unsupervised machine learning algorithm is used to detect anomalies in the historical bidding sequences of DSPs to help identify deviation behaviors; B3: Calculate the mutual information S between different DSP bidding sequences ij If S ij Exceeding the experience threshold will trigger a conspiracy alert; B4: Estimating DSP through counterfactual simulation methods i Behavior of DSP j Impact of expected returns And sum to obtain the total externality 3. The anti-fraud detection and prevention system for RTB bidding based on game theory according to claim 1, characterized in that, The Market Anti-Fraud Manager (MAM) includes a prevention module, the working logic of which is as follows: C1: Maintain and dynamically update the reputation score of each DSP. i , Historical performance rate is calculated based on the accuracy of budget expenditures, while fraud rate is calculated based on historical delta. i (t) statistics; C2: Credit score i With penalty tax rate λ and allocation weight a i (t) is associated with imposing higher penalty rates or lowering the allocation priority of low-reputation DSPs; C3: Employs reinforcement learning algorithms to ensure long-term platform benefits. k To optimize the objective, the penalty tax rate λ is dynamically updated using the following formula: Where η is the learning rate. The gradient is estimated by replaying historical logs.

4. The anti-fraud detection and prevention system for RTB bidding based on game theory according to claim 1, characterized in that, The Market Anti-Fraud Manager (MAM) includes a log auditing module, and the working logic of the log auditing module is as follows: D1: Record the request ID and the bids of each DSP at each bidding time t. i (t), calculated deviation δ i (t), penalty tax τ i (t), final payment p i (t), allocation result a i (t); D2: At the end of the settlement period, call the anti-fraud detection service to verify the authenticity of click and conversion data after the ad campaign; D3: Perform retrospective settlement based on the verification results to recover improper gains resulting from fraud or compensate for misjudgments; D4: Provides differential privacy protection for all data processing operations and provides immutability and auditability for all log records.

5. The anti-fraud detection and prevention system for RTB bidding based on game theory according to claim 1, characterized in that, The Market Anti-Fraud Manager (MAM) includes a collusion detection module, the working logic of which is as follows: E1: Based on information theory methods, compute any two DSPs i DSP j Mutual information S of bidding sequences ij ; E2: Statistical correlation of its behavior; E3: Set dynamic threshold T collusion When S ij >T collusion Time-determination DSP i DSP j There is potential for collusion: E4: Implement enhanced penalties for DSPs that trigger conspiracy alerts.

6. The anti-fraud detection and prevention system in RTB bidding based on game theory according to claim 5, characterized in that, In step S4, the punitive measures include increasing the penalty tax rate λ and temporarily freezing credit scores r. i One or more of the following measures may be taken: suspending their eligibility to participate in advertising space bidding.

7. A fraud detection and prevention system for RTB bidding based on game theory as described in claims 1-6, characterized in that, The anti-fraud detection and prevention system in RTB bidding adopts a layered computing architecture to meet the low latency requirements of RTB: Deploy lightweight detection models on edge computing nodes, utilize local caching and KNN approximation algorithms to calculate approximate values, and make real-time decisions; A high-precision Nash equilibrium simulation model is periodically run at the central computing node to update the global game model parameters and thresholds, and then distributed to the edge nodes. Edge nodes work in collaboration with central nodes to ensure that the entire system responds to bidding requests in less than 100 milliseconds.

8. A fraud detection and prevention system for RTB bidding based on game theory according to claims 1-6, characterized in that, Fraud detection and prevention methods in RTB bidding include the following steps: S1: Receive bids from multiple DSPs. i (t); S2: Based on the repeated game Nash equilibrium model, calculate the deviation δ of each bid in real time. i (t) and externality E i (t); S3: The deviation δ i (t) is compared with the dynamic threshold T, if δ i If (t) > T, then a penalty tax τ is generated. i (t), otherwise set τ i (t) = 0; S4: Calculate the final payment p i (t), and combine DPS reputation score to determine ad allocation a i (t); S5: Record all relevant data for this auction to the audit log; S6: Perform the following sub-steps periodically; S61: Estimate platform revenue R based on audit log playback. k ; S62: Update the penalty tax rate parameter λ using a reinforcement learning algorithm; S63: Perform retrospective settlement and adjust revenue distribution based on actual ad conversion data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements steps S1-S6 of the anti-fraud detection and prevention system for RTB bidding based on game theory as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements steps S1-S6 of the anti-fraud detection and prevention system for RTB bidding based on game theory as described in claim 8.

Images