Abstract:Despite ongoing efforts to defend neural classifiers from adversarial attacks, they remain vulnerable, especially to unseen attacks. In contrast, humans are difficult to be cheated by subtle manipulations, since we make judgments only based on essential factors. Inspired by this observation, we attempt to model label generation with essential label-causative factors and incorporate label-non-causative factors to assist data generation. For an adversarial example, we aim to discriminate the perturbations as non-causative factors and make predictions only based on the label-causative factors. Concretely, we propose a casual diffusion model (CausalDiff) that adapts diffusion models for conditional data generation and disentangles the two types of casual factors by learning towards a novel casual information bottleneck objective. Empirically, CausalDiff has significantly outperformed state-of-the-art defense methods on various unseen attacks, achieving an average robustness of 86.39% (+4.01%) on CIFAR-10, 56.25% (+3.13%) on CIFAR-100, and 82.62% (+4.93%) on GTSRB (German Traffic Sign Recognition Benchmark).
Abstract:Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that TGA not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good).
Abstract:Test-Time Adaptation (TTA) has emerged as a promising paradigm for enhancing the generalizability of models. However, existing mainstream TTA methods, predominantly operating at batch level, often exhibit suboptimal performance in complex real-world scenarios, particularly when confronting outliers or mixed distributions. This phenomenon stems from a pronounced over-reliance on statistical patterns over the distinct characteristics of individual instances, resulting in a divergence between the distribution captured by the model and data characteristics. To address this challenge, we propose Meet-In-The-Middle based Test-Time Adaptation ($\textbf{MITA}$), which introduces energy-based optimization to encourage mutual adaptation of the model and data from opposing directions, thereby meeting in the middle. MITA pioneers a significant departure from traditional approaches that focus solely on aligning the model to the data, facilitating a more effective bridging of the gap between model's distribution and data characteristics. Comprehensive experiments with MITA across three distinct scenarios (Outlier, Mixture, and Pure) demonstrate its superior performance over SOTA methods, highlighting its potential to significantly enhance generalizability in practical applications.
Abstract:As Large Language Models (LLMs) grow increasingly powerful, ensuring their safety and alignment with human values remains a critical challenge. Ideally, LLMs should provide informative responses while avoiding the disclosure of harmful or sensitive information. However, current alignment approaches, which rely heavily on refusal strategies, such as training models to completely reject harmful prompts or applying coarse filters are limited by their binary nature. These methods either fully deny access to information or grant it without sufficient nuance, leading to overly cautious responses or failures to detect subtle harmful content. For example, LLMs may refuse to provide basic, public information about medication due to misuse concerns. Moreover, these refusal-based methods struggle to handle mixed-content scenarios and lack the ability to adapt to context-dependent sensitivities, which can result in over-censorship of benign content. To overcome these challenges, we introduce HiddenGuard, a novel framework for fine-grained, safe generation in LLMs. HiddenGuard incorporates Prism (rePresentation Router for In-Stream Moderation), which operates alongside the LLM to enable real-time, token-level detection and redaction of harmful content by leveraging intermediate hidden states. This fine-grained approach allows for more nuanced, context-aware moderation, enabling the model to generate informative responses while selectively redacting or replacing sensitive information, rather than outright refusal. We also contribute a comprehensive dataset with token-level fine-grained annotations of potentially harmful information across diverse contexts. Our experiments demonstrate that HiddenGuard achieves over 90% in F1 score for detecting and redacting harmful content while preserving the overall utility and informativeness of the model's responses.
Abstract:The black-box nature of large language models (LLMs) poses challenges in interpreting results, impacting issues such as data intellectual property protection and hallucination tracing. Training data attribution (TDA) methods are considered effective solutions to address these challenges. Most recent TDA methods rely on influence functions, assuming the model achieves minimized empirical risk. However, achieving this criterion is difficult, and sourcing accuracy can be compromised by fitting errors during model training. In this paper, we introduce a novel TDA method called Debias and Denoise Attribution (DDA), which enhances influence functions by addressing fitting errors. Specifically, the debias strategy seeks to improve the performance of influence functions by eliminating the knowledge bias present in the base model before fine-tuning, while the denoise strategy aims to reduce discrepancies in influence scores arising from varying degrees of fitting during the training process through smoothing techniques. Experimental results demonstrate that our method significantly outperforms existing approaches, achieving an averaged AUC of 91.64%. Moreover, DDA exhibits strong generality and scalability across various sources and different-scale models like LLaMA2, QWEN2, and Mistral.
Abstract:Generative retrieval represents a novel approach to information retrieval. It uses an encoder-decoder architecture to directly produce relevant document identifiers (docids) for queries. While this method offers benefits, current approaches are limited to scenarios with binary relevance data, overlooking the potential for documents to have multi-graded relevance. Extending generative retrieval to accommodate multi-graded relevance poses challenges, including the need to reconcile likelihood probabilities for docid pairs and the possibility of multiple relevant documents sharing the same identifier. To address these challenges, we introduce a framework called GRaded Generative Retrieval (GR$^2$). GR$^2$ focuses on two key components: ensuring relevant and distinct identifiers, and implementing multi-graded constrained contrastive training. First, we create identifiers that are both semantically relevant and sufficiently distinct to represent individual documents effectively. This is achieved by jointly optimizing the relevance and distinctness of docids through a combination of docid generation and autoencoder models. Second, we incorporate information about the relationship between relevance grades to guide the training process. We use a constrained contrastive training strategy to bring the representations of queries and the identifiers of their relevant documents closer together, based on their respective relevance grades. Extensive experiments on datasets with both multi-graded and binary relevance demonstrate the effectiveness of GR$^2$.
Abstract:Recommender systems play a pivotal role in mitigating information overload in various fields. Nonetheless, the inherent openness of these systems introduces vulnerabilities, allowing attackers to insert fake users into the system's training data to skew the exposure of certain items, known as poisoning attacks. Adversarial training has emerged as a notable defense mechanism against such poisoning attacks within recommender systems. Existing adversarial training methods apply perturbations of the same magnitude across all users to enhance system robustness against attacks. Yet, in reality, we find that attacks often affect only a subset of users who are vulnerable. These perturbations of indiscriminate magnitude make it difficult to balance effective protection for vulnerable users without degrading recommendation quality for those who are not affected. To address this issue, our research delves into understanding user vulnerability. Considering that poisoning attacks pollute the training data, we note that the higher degree to which a recommender system fits users' training data correlates with an increased likelihood of users incorporating attack information, indicating their vulnerability. Leveraging these insights, we introduce the Vulnerability-aware Adversarial Training (VAT), designed to defend against poisoning attacks in recommender systems. VAT employs a novel vulnerability-aware function to estimate users' vulnerability based on the degree to which the system fits them. Guided by this estimation, VAT applies perturbations of adaptive magnitude to each user, not only reducing the success ratio of attacks but also preserving, and potentially enhancing, the quality of recommendations. Comprehensive experiments confirm VAT's superior defensive capabilities across different recommendation models and against various types of attacks.
Abstract:Retrieval-augmented generation (RAG) has emerged as a popular solution to mitigate the hallucination issues of large language models. However, existing studies on RAG seldom address the issue of predictive uncertainty, i.e., how likely it is that a RAG model's prediction is incorrect, resulting in uncontrollable risks in real-world applications. In this work, we emphasize the importance of risk control, ensuring that RAG models proactively refuse to answer questions with low confidence. Our research identifies two critical latent factors affecting RAG's confidence in its predictions: the quality of the retrieved results and the manner in which these results are utilized. To guide RAG models in assessing their own confidence based on these two latent factors, we develop a counterfactual prompting framework that induces the models to alter these factors and analyzes the effect on their answers. We also introduce a benchmarking procedure to collect answers with the option to abstain, facilitating a series of experiments. For evaluation, we introduce several risk-related metrics and the experimental results demonstrate the effectiveness of our approach.
Abstract:As the modern tool of choice for question answering, large language models (LLMs) are expected to deliver answers with up-to-date knowledge. To achieve such ideal question-answering systems, locating and then editing outdated knowledge in the natural language outputs is a general target of popular knowledge editing methods. However, this target is challenging, as both identifying which tokens to edit in the reasoning steps and ensuring the coherence of the revised reasoning chain are difficult tasks. We argue that these challenges stem from the unstructured nature of natural language outputs. To address the above challenges, we propose $\textbf{Stru}$ctural $\textbf{Edit}$ing ($\textbf{StruEdit}$), an improved baseline for knowledge editing. We first prompt LLMs to produce structured outputs consisting of reasoning triplets. Then, StruEdit removes any potentially outdated knowledge and efficiently refills the structured outputs with up-to-date information in a single step. Experimental results show that StruEdit consistently delivers the highest accuracy with lowest latency compared with other knowledge editing methods.
Abstract:Adversarial purification is one of the promising approaches to defend neural networks against adversarial attacks. Recently, methods utilizing diffusion probabilistic models have achieved great success for adversarial purification in image classification tasks. However, such methods fall into the dilemma of balancing the needs for noise removal and information preservation. This paper points out that existing adversarial purification methods based on diffusion models gradually lose sample information during the core denoising process, causing occasional label shift in subsequent classification tasks. As a remedy, we suggest to suppress such information loss by introducing guidance from the classifier confidence. Specifically, we propose Classifier-cOnfidence gUided Purification (COUP) algorithm, which purifies adversarial examples while keeping away from the classifier decision boundary. Experimental results show that COUP can achieve better adversarial robustness under strong attack methods.