Multi-modal large-model self-improved label-free harmful cause detection system and method

Abstract

The invention discloses a multi-modal large-model self-improved non-labeling harmful cause detection system and method, and relates to the technical field of operation field harmful content detection. According to the detection method of the system, explicit harmful causes easy to detect in unlabeled data are utilized to enhance the detection capability of implicit complex causes; the problems of annotation data scarcity and rapid evolution content adaptability in harmful mock detection are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of harmful content detection technology, and more specifically to a multimodal large-scale self-improving unlabeled harmful meme detection system and method. Background Technology

[0002] The development of multimodal content understanding and detection technologies has evolved from unimodal analysis to multimodal fusion and then to large-model-driven approaches. Early content detection systems mainly relied on single-modal feature extraction and classification, such as text analysis based on keyword matching or image classification based on visual features. These methods performed reasonably well when dealing with simple content, but they proved inadequate when faced with complex multimodal content.

[0003] With the rise of deep learning technology, researchers began to explore applying neural networks to content understanding tasks. However, early deep learning methods primarily targeted single data modalities, such as using convolutional neural networks (CNNs) to process images or recurrent neural networks (RNNs) to process text. This single-modal approach faces significant limitations when dealing with complex multimodal content in the real world because it cannot effectively utilize complementary information between different modalities.

[0004] To overcome this limitation, the research community has gradually moved towards multimodal learning. Early multimodal methods combined information from different modalities through simple feature concatenation or weighted fusion, but this approach often ignored the complex interactions between modalities. In recent years, attention-based multimodal learning methods have achieved significant breakthroughs, such as the successful application of the Transformer architecture in multimodal understanding tasks. The emergence of pre-trained visual-language models (VLMs) such as CLIP enables models to learn cross-modal semantic alignment on large-scale image-text pairs, thereby achieving more accurate content understanding.

[0005] In the field of harmful content detection, particularly harmful meme detection, memes, as a novel multimodal form of communication that combines images and text, are widely disseminated on social media. While memes convey humor, satire, or commentary through a combination of images and short text, they are increasingly being used as a vehicle for spreading harmful content. These harmful memes often attack the racial, gender, or nationality characteristics of individuals or groups, fueling serious discrimination and conflict. Therefore, developing effective methods to identify and combat harmful memes has become an urgent issue with significant practical implications.

[0006] Early harmful meme detection methods primarily utilized pre-trained visual-language models (VLMs) to learn cross-modal relevance and uncover deep harmful content. These methods either fine-tuned VLMs or combined frozen VLMs with learnable classifiers to detect potential harmful content. With the advent of large multimodal models (LMMs) such as GPT-4o and Qwen2.5-VL, zero-shot harmful meme detection has become possible. LMMs possess powerful multimodal reasoning capabilities, enabling them to understand and analyze complex multimodal content without any task-specific training.

[0007] However, in practical applications, existing harmful meme detection technologies face several fundamental challenges:

[0008] First, the reliance on labeled data severely limits the performance and applicability of detection systems. Existing methods heavily depend on large-scale labeled data for supervised model training or fine-tuning. However, labeling a large number of memes requires significant human resources and domain expertise, and may face challenges such as copyright restrictions. The scarcity of labeled data limits the performance and generalization ability of existing label-driven methods, as they can only learn from a limited and incomplete range of harmful meme forms. Furthermore, since meme content often involves complex cultural contexts, puns, and metaphors, labelers need considerable expertise to accurately determine whether it is harmful, further increasing the difficulty and cost of labeling.

[0009] Secondly, the rapid evolution of harmful content poses a significant challenge to the adaptability of detection systems. Harmful content and its manifestations in online media evolve rapidly, driven by breaking news or emerging events. Existing label-driven methods struggle to cope with these emerging memes because such content often exhibits distributional differences from previously seen data. More importantly, the time required for labeling and training these newly emerging memes in a short period inherently conflicts with the time-sensitive nature of harmful content. Therefore, developing more responsive detection methods capable of handling new and unseen harmful memes without relying on large amounts of labeled data remains an open and nontrivial challenge.

[0010] Third, existing methods fail to fully utilize the implicit knowledge in unlabeled data. Traditional data augmentation methods primarily expand the training set by transforming existing data (such as rotation, scaling, and cropping), but this approach has limited effectiveness in multimodal content because it cannot increase semantic diversity. While few-shot learning methods reduce reliance on large amounts of labeled data, their performance remains unsatisfactory, consistently lagging behind label-driven methods. The main limitation of these methods lies in their inability to effectively utilize the rich information contained in unlabeled memes, severely restricting their ability to bridge the gap between scarce labeled data and high detection accuracy. Furthermore, context-based few-shot learning methods are severely limited by the finite context length of LMMs, failing to provide sufficient demonstration samples for LMMs.

[0011] Fourth, existing technologies lack effective knowledge transfer mechanisms across different levels of difficulty. In the context of harmful memes, there is an important but underutilized observation: for specific harmful topics, some memes explicitly express harmful intent, making them relatively easy to detect, while others convey harm in more subtle or implicit ways, posing a greater challenge to detection. However, existing methods do not fully utilize this difference in difficulty, failing to establish a knowledge transfer mechanism from easily detectable explicit samples to difficult implicit samples. If the model can first gather experience from explicit discrimination, it can effectively help the model identify more subtle harmful information.

[0012] In summary, current harmful meme detection technologies still lack effective solutions to challenges such as scarce labeled data, adaptation to rapidly evolving content, utilization of unlabeled data, and knowledge transfer across different difficulty levels. This provides ample research space and application prospects for new technological solutions. Traditional supervised learning paradigms can no longer meet the needs of practical applications, and there is an urgent need for an innovative method that can achieve efficient detection in the absence of any labeled data and can quickly adapt to emerging harmful content.

[0013] Therefore, it is necessary to propose a self-improving, label-free harmful meme detection system and method for multimodal large models to solve the above problems. Summary of the Invention

[0014] The purpose of this invention is to solve the problems mentioned in the background art.

[0015] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0016] A multimodal, large-scale, self-improving, unlabeled harmful meme detection system, including:

[0017] The explicit meme identification module is used to process the raw unlabeled meme dataset. Based on the zero-shot detection capability and prediction confidence of the large multimodal model LMM, it identifies easily detectable explicit memes and assigns pseudo-labels. This confidence-based identification mechanism avoids the problem of manual annotation required in traditional methods and ensures that the identified explicit memes have high reliability.

[0018] The contrast pairing construction module is used to reorganize the identified explicit memes into contrast pairs, wherein the most semantically similar pseudo-benign meme is retrieved for each pseudo-harmful meme, or the most semantically similar pseudo-harmful meme is retrieved symmetrically for each pseudo-benign meme; this pairing mechanism ensures the effectiveness of contrastive learning, enabling subsequent experience collection to focus on fine-grained differential analysis.

[0019] The agent self-improvement module is used to enable the LMM agent to improve itself based on the comparison pairing, including an experience collection submodule and a reference extraction submodule;

[0020] An enhanced detection module is used to detect hidden harmful memes using a detection reference generated by the agent's self-improvement module. This module uses the refined reference set as guidance to enhance the LMM agent's ability to detect complex memes. This method not only utilizes knowledge learned from explicit samples but also rapidly adapts to emerging harmful content without any retraining.

[0021] Furthermore, the explicit meme identification module includes:

[0022] The zero-shot prediction submodule is used to utilize the multimodal reasoning capability of the LMM to perform zero-shot detection for each meme and generate prediction probabilities for the two categories of "harmful" and "benign"; these probabilities reflect the confidence of the LMM in its output.

[0023] The confidence ranking submodule is used to rank all memes according to the predicted probabilities and select the memes with the highest predicted confidence as the explicit set based on the quantile threshold; ensuring that the selected memes are those that the LMM can correctly identify with high confidence.

[0024] The pseudo-label allocation submodule is used to assign pseudo-labels to each meme in the explicit set, wherein the label category is determined by the magnitude of the predicted probabilities, thereby obtaining reliable pseudo-labeled data without manual annotation.

[0025] Furthermore, the comparison pairing construction module includes:

[0026] The feature encoding submodule is used to extract features from image and text elements of the meme using the pre-trained CLIP model's text and visual encoders, generating a unified embedding representation.

[0027] The similarity retrieval submodule is used to calculate the cross-modal similarity between each pseudo-harmful meme and all pseudo-benign memes, and retrieve the most similar pseudo-benign memes to form a comparison pair; symmetrically, it retrieves the most similar pseudo-harmful memes for each pseudo-benign meme, and finally merges them to form a complete set of comparison pairs.

[0028] Furthermore, in the intelligent agent self-improvement module:

[0029] The experience collection submodule adopts a two-step thinking chain prompting strategy to guide the LMM agent to first describe the multimodal content of the two memes in the comparison pair, and then analyze and summarize the subtle differences between the two, thereby collecting case experience.

[0030] The reference refining submodule is equipped with four atomic operations, including ADD, UPVOTE, DOWNVOTE, and EDIT, to guide the LMM agent in iteratively analyzing the experience of each case, updating and maintaining a high-level detection reference set with limited capacity, and ensuring the sufficiency and representativeness of the reference.

[0031] A self-improving, unlabeled harmful meme detection method for multimodal large models, used in the aforementioned self-improving, unlabeled harmful meme detection system for multimodal large models, specifically includes the following steps:

[0032] S1. Identify explicit memes based on confidence and segment the dataset, including using the decoder of a large multimodal model (LMM) and zero-shot detection hints to perform zero-shot prediction for each meme and obtain the predicted probability vector:

[0033] ;

[0034] in , and These represent the predicted probabilities of "harmful" and "benign," respectively.

[0035] S2. Based on the predicted probabilities generated in S1, all memes are ranked by confidence and then based on quantile thresholds. Choose the one with the highest predicted probability. Meme formation explicit set :

[0036] ;

[0037] The remaining memes constitute the latent set Assign pseudo-labels to each meme in the explicit set, the label categories being determined by... and The size relationship determines:

[0038] ;

[0039] S3. Constructing comparison pairings based on multimodal retrieval, including dividing the explicit set into pseudo-harmful modality factor sets according to pseudo-labels. and pseudo-benign model factor set :

[0040] ;

[0041] ;

[0042] For each of the aforementioned pseudo-harmful memes Retrieve the pseudo-benign model factor set The memes with the most semantic similarity are used to form a comparison pair; symmetrically, for each of the pseudo-benign memes, the memes with the most semantic similarity in the set of pseudo-harmful memes are retrieved to construct a... Ultimately, they are merged to form a complete set of contrast pairings. :

[0043] ;

[0044] S4. Experience collection for paired learning guidance, including using the LMM agent and two-step mind chain prompts. Analyze the comparison pairing set Differences between each comparison pair, and collection of case experience. :

[0045] ;

[0046] S5. Iterative reference refinement and optimization, including initializing an empty reference set. It is equipped with four atomic operations: ADD, UPVOTE, DOWNVOTE, and EDIT, to guide the LMM agent in iteratively analyzing the case experience. Each experience Updating and maintaining a reference set with limited capacity :

[0047] ;

[0048] in It is a prompt that guides the agent to update the reference set;

[0049] S6. Self-generated reference-guided latent meme detection, including for the latent set Each meme in Using the refined reference set and reasoning hints Perform detection to obtain the predicted category of the meme. :

[0050] ;

[0051] in It is a meme The prediction category.

[0052] Furthermore, in step S3, the semantic similarity function is defined as the sum of visual similarity and textual similarity:

[0053] ;

[0054] in and It is a pre-trained CLIP model for text and visual encoders.

[0055] Furthermore, the two-step thought chain prompts the LMM agent to first briefly describe the multimodal content of the two memes in the comparison pair, and then carefully analyze and summarize the subtle differences between them.

[0056] Furthermore, in step S5, the atomic operation ADD is used to discover new patterns and add new references; the UPVOTE operation is used to increase the importance of references; the DOWNVOTE operation is used to decrease the importance of references; and the EDIT operation is used to refine existing references.

[0057] Furthermore, in step S5, the capacity of the reference set is limited to L to ensure the refinement of the reference set.

[0058] Furthermore, in step S6, the self-generated reference serves as an implicit gradient signal, guiding the LMM agent to perform more accurate detection. This method requires no gradient updates or model fine-tuning.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. This invention achieves explicit meme identification in a completely unlabeled environment by utilizing the zero-sample detection capability and prediction probability of a Level Mirror (LMM) as confidence indicators. The prediction probability reflects the LMM's confidence in its output; high-confidence predictions typically correspond to easily detectable explicit memes. By setting a quantile threshold, this invention can adaptively select an appropriate number of explicit samples, avoiding problems that may arise from a fixed threshold.

[0061] 2. This invention employs multimodal retrieval technology to construct comparative pairings. By pairing memes with similar semantics but opposite labels, subsequent experience collection can focus on fine-grained difference analysis. The calculation of multimodal similarity combines visual and textual dimensions, ensuring the semantic relevance of the pairings. The pre-trained CLIP encoder effectively captures cross-modal semantic alignment.

[0062] 3. This invention effectively alleviates the illusion problem of LMM through a two-step thought chain prompting strategy. The first step requires the agent to describe the meme content and establish an accurate understanding of the input; the second step, based on this, performs differential analysis to ensure the accuracy and fine-grainedness of the experience. By analyzing a large number of contrast pairs, the agent can accumulate diverse case experience, covering various harmful expression patterns.

[0063] 4. The iterative reference refinement mechanism is one of the core innovations of this invention. Directly using a large amount of case experience would exceed the context length limit of LMM, while this invention solves this problem by refining high-level references. Four atomic operations enable the agent to flexibly maintain the reference set: the ADD operation is used to discover new patterns, the UPVOTE / DOWNVOTE operation is used to adjust the importance of references, and the EDIT operation is used to refine existing references. The capacity limit ensures the refinement of the reference set, forcing the agent to retain the most important and representative references.

[0064] 5. In this invention, a self-generated detection reference serves as an implicit "gradient signal," guiding the LMM agent to perform more accurate detection. This method requires no gradient updates or model fine-tuning, relying entirely on cue engineering and contextual learning, exhibiting exceptional flexibility and adaptability. The reference set can quickly adapt to emerging harmful content; simply re-execute the identification-pairing-experience collection-reference extraction process on new data. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the overall framework of the self-improving, unlabeled harmful meme detection method for multimodal large models of the present invention. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see Figure 1A multimodal large-scale model self-improvement unlabeled harmful meme detection system and method are presented, demonstrating the complete processing flow from inputting the original unlabeled meme dataset to the final detection result. The figure includes four main stages: (1) a confidence-based explicit meme identification stage, which divides the original dataset into explicit and implicit sets; (2) a contrast pair construction stage, which reorganizes explicit memes into contrast pairs through multimodal retrieval; (3) an agent self-improvement stage, which includes two sub-processes: experience collection and reference refinement; and (4) an enhanced detection stage, which uses self-generated references to detect implicit memes. The figure also shows the data flow and interaction relationships between the modules.

[0068] Terminology Explanation

[0069] Large Multimodal Models (LMMs) are large-scale neural network models capable of simultaneously processing and understanding multiple modalities of data (such as text, images, and videos). LMMs learn rich cross-modal knowledge and powerful reasoning capabilities through pre-training on massive amounts of multimodal data. Representative LMMs include GPT-4o, Qwen2.5-VL, and LLaVA. In this invention, the LMM is used as the core of the agent, responsible for tasks such as zero-shot prediction, experience collection, reference extraction, and final detection. Its powerful multimodal understanding and reasoning capabilities provide crucial support for this invention.

[0070] The CLIP encoder is a multimodal pre-trained encoder model developed by OpenAI, based on contrastive learning. CLIP (Contrastive Language-Image Pre-training) learns the semantic correspondence between visual and textual modalities through contrastive learning training on large-scale image-text pair data. The CLIP encoder consists of a visual encoder and a text encoder, used to extract image features and text features respectively, and map them to the same feature space. In this invention, the CLIP encoder is used to calculate multimodal similarity, providing reliable semantic matching capabilities for constructing contrastive pairings.

[0071] Chain-of-Thought Prompting (CoT) is a prompting technique that guides large language models through step-by-step reasoning. CoT significantly improves model performance on complex tasks and reduces illusion problems by requiring the model to demonstrate its reasoning process before providing the final answer. In the experience collection phase of this invention, a two-step CoT strategy is employed: the first step requires the agent to describe the meme content, and the second step requires the agent to analyze discrepancies. This design ensures the agent's accurate understanding of the input and enables fine-grained discrepancy analysis.

[0072] Zero-shot learning refers to the ability of a model to complete a task without having seen any training samples for that specific task, relying solely on pre-trained knowledge and the task description. In this invention, the zero-shot detection capability of LMM is used to identify explicit memes, eliminating the need for labeled data for harmful meme detection and significantly reducing the system's data dependency.

[0073] Experimental results

[0074] 1. Experimental setup

[0075] 1.1 Dataset

[0076] To fully verify the effectiveness and generalization ability of this invention, experimental evaluations were conducted on three public and diverse benchmark datasets:

[0077] The FHM dataset is a widely adopted dataset for detecting harmful memes in English, containing 9,000 samples, of which 3,300 are harmful and 5,700 are benign. The dataset is divided into 8,500 training samples and 500 test samples.

[0078] MAMI dataset: Another important benchmark for detecting harmful English memes, containing 11,000 samples, half harmful and half benign (5,500 each). The dataset includes 10,000 training samples and 1,000 test samples.

[0079] ToxiCN dataset: A Chinese harmful meme detection dataset containing 12,000 samples, of which 3,827 are harmful and 8,173 are benign. The dataset is divided into 9,600 training samples and 2,400 test samples.

[0080] These three datasets not only cover different languages ​​(English and Chinese), but also represent different types of harmful content and expressions, which can fully verify the cross-language and cross-scenario applicability of the present invention.

[0081] 1.2 Evaluation Indicators

[0082] Following previous work in the field of harmful meme detection, we employ accuracy and macro-F1 score as evaluation metrics. Accuracy reflects the overall classification correctness of the model, while the macro-F1 score comprehensively considers precision and recall, and is particularly suitable for class imbalance scenarios. The combined use of these two metrics allows for a comprehensive evaluation of the model's detection performance.

[0083] 1.3 Baseline Model

[0084] To comprehensively evaluate the performance of this invention (ALARM), it was compared with fourteen representative benchmark models, which can be divided into two groups:

[0085] (1) Label-driven methods: This group of methods requires labeled training data to optimize the detection model, including: - MOMENTA: a multimodal hate content detection method based on momentum contrast learning - PromptHate: a method for harmful meme detection using cue learning - MR.HARM: a multimodal relation-aware harmful meme detection method - Pro-Cap: a harmful meme detection method based on question-and-answer image description - ISM: a harmful meme detection method based on invariant semantic modeling - ExplainHM: an interpretable harmful meme detection method - HHPrompt: a harmful meme detection method based on hierarchical cue learning.

[0086] (2) Few-shot learning methods: This group of methods uses only a small number of labeled samples to enhance detection capabilities, including:

[0087] OPT-30B: A large-scale language model that learns from few-shot contexts;

[0088] OpenFlamingo-9B: An open-source multimodal model that supports few-shot learning;

[0089] Qwen2.5-VL-72B: A large-scale vision-language model that supports few-shot context learning;

[0090] GPT-4o: OpenAI's large-scale multimodal model, demonstrated using few samples;

[0091] Mod-HATE: A modular harmful content detection method trained on a small number of samples;

[0092] LoReHM: A low-resource harmful meme detection method that uses contextual learning;

[0093] It is important to note that, except for Mod-HATE which is trained on a small number of labeled samples, all other few-shot methods use samples as contextual examples. Following the setup of previous work, 50 example samples are provided for the few-shot baseline.

[0094] 1.4 Implementation Details

[0095] This invention primarily uses the open-source Qwen2.5-VL-72B model as the LMM backbone. Given the high scalability of this invention, the closed-source GPT-4o (gpt-4o-2024-11-20 version) was also selected as another backbone for verification. However, the confidence-based explicit meme identification mechanism still requires the use of the open-source model Qwen2.5-VL-72B because it needs access to the predicted probabilities.

[0096] In the multimodal retrieval phase, a pre-trained CLIP model with text and visual encoders is used. Confidence selection ratio. Set to 0.2 on all three datasets, reference set size Set to 15. Although this invention can directly process the test set for reference extraction, to ensure fair comparison, this invention is limited to using only the training set to generate the reference.

[0097] The experiment was conducted on an NVIDIA L40s GPU.

[0098] 2. Main performance results

[0099] Under the same experimental setup, ALARM was compared with all benchmark models, and the experimental results are shown in the table below:

[0100] Based on the above experimental results, the following important observations were made:

[0101] (1) The ALARM method of this invention achieves superior detection performance, even surpassing label-driven methods. ALARM achieves optimal or near-optimal performance on three diverse benchmark datasets. Specifically, ALARM using Qwen2.5-VL-72B as its backbone achieves 75.80% accuracy and 75.79% macro F1 score on the FHM dataset, representing improvements of 0.20% and 0.40% respectively compared to the best label-driven baseline, ExplainHM; and 79.21% accuracy and 72.51% macro F1 score on the ToxiCN dataset, representing improvements of 1.34% and 2.26% respectively compared to the best label-driven baseline, MOMENTA. ALARM using GPT-4o as its backbone achieves 85.50% accuracy and macro F1 score on the MAMI dataset, significantly outperforming all baseline methods.

[0102] More importantly, ALARM significantly enhances the detection capability of its LMM backbone. Compared to the original Qwen2.5-VL-72B zero-sample performance, ALARM achieves an average accuracy improvement of 6.35%. This result fully demonstrates that the present invention can effectively improve the detection capability of complex latent samples in completely unlabeled data by making full use of easily detectable explicit samples in unlabeled data.

[0103] (2) Label-driven methods perform well on specific datasets but have significant limitations. These methods primarily rely on carefully designed supervised learning strategies to optimize on the corresponding training data. For example, Pro-Cap uses multimodal content and question-answering-based image descriptions to fine-tune the model to improve detection accuracy. However, they require a large amount of labeled data to achieve competitive performance, which is often impractical in real-world scenarios due to the labor-intensive nature of labeling and potential copyright restrictions. Furthermore, label-driven methods tend to specifically capture harmful patterns on specific datasets, which limits their ability to effectively generalize to emerging, unseen memes, making them unable to cope with rapidly evolving harmful memes.

[0104] (3) Although few-shot learning methods reduce the reliance on large amounts of labeled data, their performance remains unsatisfactory. These methods consistently lag behind label-driven methods, primarily due to their inability to effectively utilize the rich information contained in unlabeled memes. This severely limits their ability to bridge the gap between scarce labeled data and high detection accuracy. Furthermore, context-based few-shot learning methods are severely constrained by the limited context length of LMMs, failing to provide sufficient demonstration samples for LMMs. This invention effectively alleviates this challenge through a reference extraction mechanism, extracting only a small number of high-level references to well represent diverse and rich meme cases.

[0105] (4) This invention has the unique advantage of being completely unannotated. Unlike all baseline methods, ALARM achieves excellent performance without any manually annotated data, which allows it to quickly adapt to emerging forms and topics of harmful content without any additional annotation work. This feature is of great significance in practical applications because harmful memes evolve much faster than the speed at which annotated data can be obtained.

[0106] 3. Ablation test

[0107] To analyze the role of each core component in this invention, a comprehensive ablation study was conducted, and the results are shown in the table below:

[0108] variants FHM MAMI ToxiCN Acc M-F1 Acc M-F1 Acc M-F1 w / o Confidence 70.78 70.79 79.90 79.09 75.37 68.49 w / o Pairing 72.40 72.33 78.70 78.69 74.66 67.13 w / o Experience 72.00 71.82 80.00 79.98 76.99 70.75 w / o Reference 69.60 68.46 78.10 77.90 73.25 65.84 ALARM 75.80 75.79 81.28 81.25 79.21 72.51

[0109] (1) Effectiveness of the confidence-based explicit meme identification mechanism. A variant model "w / oConfidence" was designed, which replaces the original probability-based mechanism with a simple random selection strategy, randomly selecting... Memes were used as explicit samples and assigned pseudo-labels. As can be seen from the table, this variant exhibits a significant performance drop because random selection inevitably includes complex and subtle memes. These memes are more likely to be assigned incorrect pseudo-labels by the LMM without any task-specific fine-tuning, thus introducing significant noise into subsequent processes and leading to performance degradation. This result validates the importance of confidence-based recognition mechanisms.

[0110] (2) Contribution of the contrastive pairing mechanism. The “w / o Pairing” variant completely removes the retrieval-based contrastive pairing mechanism, randomly realigning the memes from the two pseudo-classes in the original pairing to form a new pairing. Semantically mismatched paired memes undermine the core principle of contrastive learning, reduce the quality of the collected experience, and lead to a decline in detection performance.

[0111] (3) The necessity of the experience collection process. The “w / o Experience” variant excludes the experience collection process and extracts the reference directly from the initial comparison pair. Experimental results show that the detection reference extracted directly from the paired memes is of poor quality, provides limited guidance for the LMM agent, and is difficult to detect more complex memes.

[0112] (4) The key role of the reference extraction mechanism. The “w / o Reference” variant eliminates the reference extraction process and directly utilizes experience to enhance the detection capabilities of the LMM agent. Experimental results show that the collected experience is still case-specific and cannot provide more generalized and higher-level guidance for robust detection. In addition, due to the context length limitation, only a small amount of experience can be input into the LMM, which also limits the effectiveness of this variant.

[0113] The above ablation experiments fully verify the effectiveness and necessity of each core component of the present invention.

[0114] In summary, the experimental results fully verify the effectiveness and superiority of the self-improving, label-free harmful meme detection method based on large-scale multimodal model agents proposed in this invention. It provides a feasible and efficient solution to problems such as scarce labeled data and rapid content adaptation in practical applications. This invention not only achieves or even surpasses label-driven methods in detection performance, but more importantly, it realizes a completely label-free detection paradigm, possessing strong practical value and application prospects.

[0115] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the content of the present invention's specification shall also be included within the scope of protection of the present invention.

Claims

1. A self-improving, label-free harmful meme detection system for multimodal large models, characterized in that, include: The explicit meme identification module is used to process the raw unlabeled meme dataset. Based on the zero-shot detection capability and prediction confidence of the large multimodal model LMM, it identifies easily detectable explicit memes and assigns pseudo-labels. The comparison pairing construction module is used to reorganize the identified explicit memes into comparison pairs, wherein the most semantically similar pseudo-benign meme is retrieved for each pseudo-harmful meme, or the most semantically similar pseudo-harmful meme is retrieved symmetrically for each pseudo-benign meme. The agent self-improvement module is used to enable the LMM agent to improve itself based on the comparison pairing, including an experience collection submodule and a reference extraction submodule; An enhanced detection module is used to detect hidden harmful memes using a detection reference generated by the agent's self-improvement module.

2. The multimodal large model self-improvement unlabeled harmful meme detection system according to claim 1, characterized in that: The explicit meme identification module includes: The zero-shot prediction submodule is used to utilize the multimodal reasoning capability of the LMM to perform zero-shot detection on each meme and generate prediction probabilities for the two categories of "harmful" and "benign". The confidence ranking submodule is used to rank all memes according to the predicted probabilities and select the memes with the highest predicted confidence as the explicit set based on the quantile threshold. The pseudo-label allocation submodule is used to assign pseudo-labels to each meme in the explicit set, wherein the label category is determined by the magnitude relationship of the predicted probabilities.

3. The multimodal large model self-improvement unlabeled harmful meme detection system according to claim 1, characterized in that: The comparison and pairing construction module includes: The feature encoding submodule is used to extract features from image and text elements of the meme using the pre-trained CLIP model's text and visual encoders, generating a unified embedding representation. The similarity retrieval submodule is used to calculate the cross-modal similarity between each pseudo-harmful meme and all pseudo-benign memes, and retrieve the most similar pseudo-benign memes to form a comparison pair; symmetrically, it retrieves the most similar pseudo-harmful memes for each pseudo-benign meme, and finally merges them to form a complete set of comparison pairs.

4. The multimodal large model self-improvement unlabeled harmful meme detection system according to claim 1, characterized in that: In the agent self-improvement module: The experience collection submodule adopts a two-step thinking chain prompting strategy to guide the LMM agent to first describe the multimodal content of the two memes in the comparison pair, and then analyze and summarize the subtle differences between the two, thereby collecting case experience. The reference refining submodule is equipped with four atomic operations, including ADD, UPVOTE, DOWNVOTE, and EDIT, to guide the LMM agent in iteratively analyzing the experience of each case and updating and maintaining the high-level detection reference set with limited capacity.

5. A self-improving, unlabeled harmful meme detection method for multimodal large models, used in the self-improving, unlabeled harmful meme detection system for multimodal large models according to any one of claims 1-4, characterized in that, Specifically, the following steps are included: S1. Identify explicit memes based on confidence and segment the dataset, including using the decoder of a large multimodal model (LMM) and zero-shot detection hints to perform zero-shot prediction for each meme and obtain the predicted probability vector: ； in , and These represent the predicted probabilities of "harmful" and "benign," respectively. S2. Based on the predicted probabilities generated in S1, all memes are ranked by confidence and then based on quantile thresholds. Choose the one with the highest predicted probability. Meme formation explicit set : ； The remaining memes constitute the latent set Assign pseudo-labels to each meme in the explicit set, the label categories being determined by... and The size relationship determines: ； S3. Constructing comparison pairings based on multimodal retrieval, including dividing the explicit set into pseudo-harmful modality factor sets according to pseudo-labels. and pseudo-benign model factor set : ； ； For each of the aforementioned pseudo-harmful memes Retrieve the pseudo-benign model factor set The memes with the most semantic similarity are used to form a comparison pair; symmetrically, for each of the pseudo-benign memes, the memes with the most semantic similarity in the set of pseudo-harmful memes are retrieved to construct a... Ultimately, they are merged to form a complete set of contrast pairings. : ； S4. Experience collection for paired learning guidance, including using the LMM agent and two-step mind chain prompts. Analyze the comparison pairing set Differences between each comparison pair, and collection of case experience. : ； S5. Iterative reference refinement and optimization, including initializing an empty reference set. It is equipped with four atomic operations: ADD, UPVOTE, DOWNVOTE, and EDIT, to guide the LMM agent in iteratively analyzing the case experience. Each experience Updating and maintaining a reference set with limited capacity : ； in It is a prompt that guides the agent to update the reference set; S6. Self-generated reference-guided latent meme detection, including for the latent set Each meme in Using the refined reference set and reasoning hints Perform detection to obtain the predicted category of the meme. : ； in It is a meme The prediction category.

6. The multimodal large model self-improvement unlabeled harmful meme detection system and method according to claim 5, characterized in that: In step S3, the semantic similarity function is defined as the sum of visual similarity and textual similarity: ； in and It is a pre-trained CLIP model for text and visual encoders.

7. The method for detecting unlabeled harmful memes in a multimodal large model with self-improvement according to claim 5, characterized in that: The two-step thought chain prompts the LMM agent to first briefly describe the multimodal content of the two memes in the comparison pair, and then carefully analyze and summarize the subtle differences between them.

8. The method for detecting unlabeled harmful memes in a multimodal large model with self-improvement according to claim 5, characterized in that: In step S5, the atomic operation ADD is used to discover new patterns and add new references; the UPVOTE operation is used to increase the importance of references; the DOWNVOTE operation is used to decrease the importance of references; and the EDIT operation is used to refine existing references.

9. The method for detecting unlabeled harmful memes in a multimodal large model with self-improvement according to claim 5, characterized in that: In step S5, the capacity of the reference set is limited to L to ensure the refinement of the reference set.

10. The method for detecting unlabeled harmful memes in a multimodal large model with self-improvement according to claim 5, characterized in that: In step S6, the self-generated reference serves as an implicit gradient signal, guiding the LMM agent to perform more accurate detection. This method requires no gradient updates or model fine-tuning.

Images