Reinforcement Learning for Self-Improving Agent with Skill Library

Large Language Model (LLM)-based agents have demonstrated remarkable capabilities in complex reasoning and multi-turn interactions but struggle to continuously improve and adapt when deployed in new environments. One promising approach is implementing skill libraries that allow agents to learn, validate, and apply new skills. However, current skill library approaches rely primarily on LLM prompting, making consistent skill library implementation challenging. To overcome these challenges, we propose a Reinforcement Learning (RL)-based approach to enhance agents' self-improvement capabilities with a skill library. Specifically, we introduce Skill Augmented GRPO for self-Evolution (SAGE), a novel RL framework that systematically incorporates skills into learning. The framework's key component, Sequential Rollout, iteratively deploys agents across a chain of similar tasks for each rollout. As agents navigate through the task chain, skills generated from previous tasks accumulate in the library and become available for subsequent tasks. Additionally, the framework enhances skill generation and utilization through a Skill-integrated Reward that complements the original outcome-based rewards. Experimental results on AppWorld demonstrate that SAGE, when applied to supervised-finetuned model with expert experience, achieves 8.9% higher Scenario Goal Completion while requiring 26% fewer interaction steps and generating 59% fewer tokens, substantially outperforming existing approaches in both accuracy and efficiency.

Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset

Machine unlearning has emerged as an effective strategy for forgetting specific information in the training data. However, with the increasing integration of visual data, privacy concerns in Vision Language Models (VLMs) remain underexplored. To address this, we introduce Facial Identity Unlearning Benchmark (FIUBench), a novel VLM unlearning benchmark designed to robustly evaluate the effectiveness of unlearning algorithms under the Right to be Forgotten setting. Specifically, we formulate the VLM unlearning task via constructing the Fictitious Facial Identity VQA dataset and apply a two-stage evaluation pipeline that is designed to precisely control the sources of information and their exposure levels. In terms of evaluation, since VLM supports various forms of ways to ask questions with the same semantic meaning, we also provide robust evaluation metrics including membership inference attacks and carefully designed adversarial privacy attacks to evaluate the performance of algorithms. Through the evaluation of four baseline VLM unlearning algorithms within FIUBench, we find that all methods remain limited in their unlearning performance, with significant trade-offs between model utility and forget quality. Furthermore, our findings also highlight the importance of privacy attacks for robust evaluations. We hope FIUBench will drive progress in developing more effective VLM unlearning algorithms.

FATH: Authentication-based Test-time Defense against Indirect Prompt Injection Attacks

Large language models (LLMs) have been widely deployed as the backbone with additional tools and text information for real-world applications. However, integrating external information into LLM-integrated applications raises significant security concerns. Among these, prompt injection attacks are particularly threatening, where malicious instructions injected in the external text information can exploit LLMs to generate answers as the attackers desire. While both training-time and test-time defense methods have been developed to mitigate such attacks, the unaffordable training costs associated with training-time methods and the limited effectiveness of existing test-time methods make them impractical. This paper introduces a novel test-time defense strategy, named Formatting AuThentication with Hash-based tags (FATH). Unlike existing approaches that prevent LLMs from answering additional instructions in external text, our method implements an authentication system, requiring LLMs to answer all received instructions with a security policy and selectively filter out responses to user instructions as the final output. To achieve this, we utilize hash-based authentication tags to label each response, facilitating accurate identification of responses according to the user's instructions and improving the robustness against adaptive attacks. Comprehensive experiments demonstrate that our defense method can effectively defend against indirect prompt injection attacks, achieving state-of-the-art performance under Llama3 and GPT3.5 models across various attack methods. Our code is released at: https://github.com/Jayfeather1024/FATH

Consistency Purification: Effective and Efficient Diffusion Purification towards Certified Robustness

Diffusion Purification, purifying noised images with diffusion models, has been widely used for enhancing certified robustness via randomized smoothing. However, existing frameworks often grapple with the balance between efficiency and effectiveness. While the Denoising Diffusion Probabilistic Model (DDPM) offers an efficient single-step purification, it falls short in ensuring purified images reside on the data manifold. Conversely, the Stochastic Diffusion Model effectively places purified images on the data manifold but demands solving cumbersome stochastic differential equations, while its derivative, the Probability Flow Ordinary Differential Equation (PF-ODE), though solving simpler ordinary differential equations, still requires multiple computational steps. In this work, we demonstrated that an ideal purification pipeline should generate the purified images on the data manifold that are as much semantically aligned to the original images for effectiveness in one step for efficiency. Therefore, we introduced Consistency Purification, an efficiency-effectiveness Pareto superior purifier compared to the previous work. Consistency Purification employs the consistency model, a one-step generative model distilled from PF-ODE, thus can generate on-manifold purified images with a single network evaluation. However, the consistency model is designed not for purification thus it does not inherently ensure semantic alignment between purified and original images. To resolve this issue, we further refine it through Consistency Fine-tuning with LPIPS loss, which enables more aligned semantic meaning while keeping the purified images on data manifold. Our comprehensive experiments demonstrate that our Consistency Purification framework achieves state-of the-art certified robustness and efficiency compared to baseline methods.

Safeguarding Vision-Language Models Against Patched Visual Prompt Injectors

Large language models have become increasingly prominent, also signaling a shift towards multimodality as the next frontier in artificial intelligence, where their embeddings are harnessed as prompts to generate textual content. Vision-language models (VLMs) stand at the forefront of this advancement, offering innovative ways to combine visual and textual data for enhanced understanding and interaction. However, this integration also enlarges the attack surface. Patch-based adversarial attack is considered the most realistic threat model in physical vision applications, as demonstrated in many existing literature. In this paper, we propose to address patched visual prompt injection, where adversaries exploit adversarial patches to generate target content in VLMs. Our investigation reveals that patched adversarial prompts exhibit sensitivity to pixel-wise randomization, a trait that remains robust even against adaptive attacks designed to counteract such defenses. Leveraging this insight, we introduce SmoothVLM, a defense mechanism rooted in smoothing techniques, specifically tailored to protect VLMs from the threat of patched visual prompt injectors. Our framework significantly lowers the attack success rate to a range between 0% and 5.0% on two leading VLMs, while achieving around 67.3% to 95.0% context recovery of the benign images, demonstrating a balance between security and usability.

Mitigating Fine-tuning Jailbreak Attack with Backdoor Enhanced Alignment

Despite the general capabilities of Large Language Models (LLMs) like GPT-4 and Llama-2, these models still request fine-tuning or adaptation with customized data when it comes to meeting the specific business demands and intricacies of tailored use cases. However, this process inevitably introduces new safety threats, particularly against the Fine-tuning based Jailbreak Attack (FJAttack), where incorporating just a few harmful examples into the fine-tuning dataset can significantly compromise the model safety. Though potential defenses have been proposed by incorporating safety examples into the fine-tuning dataset to reduce the safety issues, such approaches require incorporating a substantial amount of safety examples, making it inefficient. To effectively defend against the FJAttack with limited safety examples, we propose a Backdoor Enhanced Safety Alignment method inspired by an analogy with the concept of backdoor attacks. In particular, we construct prefixed safety examples by integrating a secret prompt, acting as a "backdoor trigger", that is prefixed to safety examples. Our comprehensive experiments demonstrate that through the Backdoor Enhanced Safety Alignment with adding as few as 11 prefixed safety examples, the maliciously fine-tuned LLMs will achieve similar safety performance as the original aligned models. Furthermore, we also explore the effectiveness of our method in a more practical setting where the fine-tuning data consists of both FJAttack examples and the fine-tuning task data. Our method shows great efficacy in defending against FJAttack without harming the performance of fine-tuning tasks.

Preference Poisoning Attacks on Reward Model Learning

Learning utility, or reward, models from pairwise comparisons is a fundamental component in a number of application domains. These approaches inherently entail collecting preference information from people, with feedback often provided anonymously. Since preferences are subjective, there is no gold standard to compare against; yet, reliance of high-impact systems on preference learning creates a strong motivation for malicious actors to skew data collected in this fashion to their ends. We investigate the nature and extent of this vulnerability systematically by considering a threat model in which an attacker can flip a small subset of preference comparisons with the goal of either promoting or demoting a target outcome. First, we propose two classes of algorithmic approaches for these attacks: a principled gradient-based framework, and several variants of rank-by-distance methods. Next, we demonstrate the efficacy of best attacks in both these classes in successfully achieving malicious goals on datasets from three diverse domains: autonomous control, recommendation system, and textual prompt-response preference learning. We find that the best attacks are often highly successful, achieving in the most extreme case 100% success rate with only 0.3% of the data poisoned. However, which attack is best can vary significantly across domains, demonstrating the value of our comprehensive vulnerability analysis that involves several classes of attack algorithms. In addition, we observe that the simpler and more scalable rank-by-distance approaches are often competitive with the best, and on occasion significantly outperform gradient-based methods. Finally, we show that several state-of-the-art defenses against other classes of poisoning attacks exhibit, at best, limited efficacy in our setting.

Test-time Backdoor Mitigation for Black-Box Large Language Models with Defensive Demonstrations

Existing studies in backdoor defense have predominantly focused on the training phase, overlooking the critical aspect of testing time defense. This gap becomes particularly pronounced in the context of Large Language Models (LLMs) deployed as Web Services, which typically offer only black-box access, rendering training-time defenses impractical. To bridge this gap, our work introduces defensive demonstrations, an innovative backdoor defense strategy for blackbox large language models. Our method involves identifying the task and retrieving task-relevant demonstrations from an uncontaminated pool. These demonstrations are then combined with user queries and presented to the model during testing, without requiring any modifications/tuning to the black-box model or insights into its internal mechanisms. Defensive demonstrations are designed to counteract the adverse effects of triggers, aiming to recalibrate and correct the behavior of poisoned models during test-time evaluations. Extensive experiments show that defensive demonstrations are effective in defending both instance-level and instruction-level backdoor attacks, not only rectifying the behavior of poisoned models but also surpassing existing baselines in most scenarios.

On the Exploitability of Reinforcement Learning with Human Feedback for Large Language Models

Reinforcement Learning with Human Feedback (RLHF) is a methodology designed to align Large Language Models (LLMs) with human preferences, playing an important role in LLMs alignment. Despite its advantages, RLHF relies on human annotators to rank the text, which can introduce potential security vulnerabilities if any adversarial annotator (i.e., attackers) manipulates the ranking score by up-ranking any malicious text to steer the LLM adversarially. To assess the red-teaming of RLHF against human preference data poisoning, we propose RankPoison, a poisoning attack method on candidates' selection of preference rank flipping to reach certain malicious behaviors (e.g., generating longer sequences, which can increase the computational cost). With poisoned dataset generated by RankPoison, we can perform poisoning attacks on LLMs to generate longer tokens without hurting the original safety alignment performance. Moreover, applying RankPoison, we also successfully implement a backdoor attack where LLMs can generate longer answers under questions with the trigger word. Our findings highlight critical security challenges in RLHF, underscoring the necessity for more robust alignment methods for LLMs.

On the Exploitability of Instruction Tuning

Instruction tuning is an effective technique to align large language models (LLMs) with human intents. In this work, we investigate how an adversary can exploit instruction tuning by injecting specific instruction-following examples into the training data that intentionally changes the model's behavior. For example, an adversary can achieve content injection by injecting training examples that mention target content and eliciting such behavior from downstream models. To achieve this goal, we propose \textit{AutoPoison}, an automated data poisoning pipeline. It naturally and coherently incorporates versatile attack goals into poisoned data with the help of an oracle LLM. We showcase two example attacks: content injection and over-refusal attacks, each aiming to induce a specific exploitable behavior. We quantify and benchmark the strength and the stealthiness of our data poisoning scheme. Our results show that AutoPoison allows an adversary to change a model's behavior by poisoning only a small fraction of data while maintaining a high level of stealthiness in the poisoned examples. We hope our work sheds light on how data quality affects the behavior of instruction-tuned models and raises awareness of the importance of data quality for responsible deployments of LLMs. Code is available at \url{https://github.com/azshue/AutoPoison}.