ICLR 2025 论文列表

Large language models (LLMs) are being rapidly developed, and a key component of their widespread deployment is their safety-related alignment. Many red-teaming efforts aim to jailbreak LLMs, where among these efforts, the Greedy Coordinate Gradient (GCG) attack's success has led to a growing interest in the study of optimization-based jailbreaking techniques. Although GCG is a significant milestone, its attacking efficiency remains unsatisfactory. In this paper, we present several improved (empirical) techniques for optimization-based jailbreaks like GCG. We first observe that the single target template of ”Sure'' largely limits the attacking performance of GCG; given this, we propose to apply diverse target templates containing harmful self-suggestion and/or guidance to mislead LLMs. Besides, from the optimization aspects, we propose an automatic multi-coordinate updating strategy in GCG (i.e., adaptively deciding how many tokens to replace in each step) to accelerate convergence, as well as tricks like easy-to-hard initialization. Then, we combine these improved technologies to develop an efficient jailbreak method, dubbed $\mathcal{I}$-GCG. In our experiments, we evaluate our $\mathcal{I}$-GCG on a series of benchmarks (such as NeurIPS 2023 Red Teaming Track). The results demonstrate that our improved techniques can help GCG outperform state-of-the-art jailbreaking attacks and achieve a nearly 100\% attack success rate. The code is released at https://github.com/jiaxiaojunQAQ/I-GCG.

Jailbreaking attacks can effectively manipulate open-source large language models (LLMs) to produce harmful responses. However, these attacks exhibit limited transferability, failing to disrupt proprietary LLMs consistently. To reliably identify vulnerabilities in proprietary LLMs, this work investigates the transferability of jailbreaking attacks by analysing their impact on the model's intent perception. By incorporating adversarial sequences, these attacks can redirect the source LLM's focus away from malicious-intent tokens in the original input, thereby obstructing the model's intent recognition and eliciting harmful responses. Nevertheless, these adversarial sequences fail to mislead the target LLM's intent perception, allowing the target LLM to refocus on malicious-intent tokens and abstain from responding. Our analysis further reveals the inherent $\textit{distributional dependency}$ within the generated adversarial sequences, whose effectiveness stems from overfitting the source LLM's parameters, resulting in limited transferability to target LLMs. To this end, we propose the Perceived-importance Flatten (PiF) method, which uniformly disperses the model's focus across neutral-intent tokens in the original input, thus obscuring malicious-intent tokens without relying on overfitted adversarial sequences. Extensive experiments demonstrate that PiF provides an effective and efficient red-teaming evaluation for proprietary LLMs.

Diffusion models build a new milestone for image generation yet raising public concerns, for they can be fine-tuned on unauthorized images for customization. Protection based on adversarial attacks rises to encounter this unauthorized diffusion customization, by adding protective watermarks to images and poisoning diffusion models. However, current protection, leveraging untargeted attacks, does not appear to be effective enough. In this paper, we propose a simple yet effective improvement for the protection against unauthorized diffusion customization by introducing targeted attacks. We show that by carefully selecting the target, targeted attacks significantly outperform untargeted attacks in poisoning diffusion models and degrading the customization image quality. Extensive experiments validate the superiority of our method on two mainstream customization methods of diffusion models, compared to existing protections. To explain the surprising success of targeted attacks, we delve into the mechanism of attack-based protections and propose a hypothesis based on our observation, which enhances the comprehension of attack-based protections. To the best of our knowledge, we are the first to both reveal the vulnerability of diffusion models to targeted attacks and leverage targeted attacks to enhance protection against unauthorized diffusion customization.

Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., *"How to make a Molotov cocktail?"* to *"How did people make a Molotov cocktail?"*) is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o-mini, GPT-4o, o1-mini, o1-preview, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1\% using direct requests to 88\% using 20 past-tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques---such as SFT, RLHF, and adversarial training---employed to align the studied models can be brittle and do not always generalize as intended. We provide code and jailbreak artifacts at https://github.com/tml-epfl/llm-past-tense.

Similar to other machine learning frameworks, Offline Reinforcement Learning (RL) is shown to be vulnerable to poisoning attacks, due to its reliance on externally sourced datasets, a vulnerability that is exacerbated by its sequential nature. To mitigate the risks posed by RL poisoning, we extend certified defenses to provide larger guarantees against adversarial manipulation, ensuring robustness for both per-state actions, and the overall expected cumulative reward. Our approach leverages properties of Differential Privacy, in a manner that allows this work to span both continuous and discrete spaces, as well as stochastic and deterministic environments---significantly expanding the scope and applicability of achievable guarantees. Empirical evaluations demonstrate that our approach ensures the performance drops to no more than 50% with up to 7% of the training data poisoned, significantly improving over the 0.008% in prior work (Wu et al., 2022), while producing certified radii that is 5 times larger as well. This highlights the potential of our framework to enhance safety and reliability in offline RL.

Transferability of adversarial examples is a well-known property that endangers all classification models, even those that are only accessible through black-box queries. Prior work has shown that an ensemble of models is more resilient to transferability: the probability that an adversarial example is effective against most models of the ensemble is low. Thus, most ongoing research focuses on improving ensemble diversity. Another line of prior work has shown that Lipschitz continuity of the models can make models more robust since it limits how a model's output changes with small input perturbations. {\em In this paper, we study the effect of Lipschitz continuity on transferability rates.} We show that although a lower Lipschitz constant increases the robustness of a single model, it is not as beneficial in training robust ensembles as it increases the transferability rate of adversarial examples across models in the ensemble. Therefore, we introduce LOTOS, a new training paradigm for ensembles, which counteracts this adverse effect. It does so by promoting orthogonality among the top-$k$ sub-spaces of the transformations of the corresponding affine layers of any pair of models in the ensemble. We theoretically show that $k$ does not need to be large for convolutional layers, which makes the computational overhead negligible. Through various experiments, we show LOTOS increases the robust accuracy of ensembles of ResNet-18 models by $6$ percentage points (p.p) against black-box attacks on CIFAR-10. It is also capable of combining with the robustness of prior state-of-the-art methods for training robust ensembles to enhance their robust accuracy by $10.7$ p.p.

Best-of-N (BoN) is a popular and effective algorithm for aligning language models to human preferences. The algorithm works as follows: at inference time, N samples are drawn from the language model, and the sample with the highest reward, as judged by a reward model, is returned as the output. Despite its effectiveness, BoN is computationally expensive; it reduces sampling throughput by a factor of N. To make BoN more efficient at inference time, one strategy is to fine-tune the language model to mimic what BoN does during inference. To achieve this, we derive the distribution induced by the BoN algorithm. We then propose to fine-tune the language model to minimize backward KL divergence to the BoN distribution. Our approach is analogous to mean-field variational inference and, thus, we term it variational BoN (vBoN). To the extent this fine-tuning is successful and we end up with a good approximation, we have reduced the inference cost by a factor of N. Our experiments on controlled generation and summarization tasks show that BoN is the most effective alignment method, and our variational approximation to BoN achieves the closest performance to BoN and surpasses models fine-tuned using the standard KL-constrained RL objective. In the controlled generation task, vBoN appears more frequently on the Pareto frontier of reward and KL divergence compared to other alignment methods. In the summarization task, vBoN achieves high reward values across various sampling temperatures.

Recent self-rewarding large language models (LLM) have successfully applied LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technologies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee of accuracy in the rewarding and ranking, which is critical for ensuring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning framework for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that leverages the consistency of rewards across different iterations to regularize the self-rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superiority of CREAM in improving both reward consistency and alignment performance. The code is publicly available at https://github.com/Raibows/CREAM.

This work studies the task of poisoned sample detection for defending against data poisoning based backdoor attacks. Its core challenge is finding a generalizable and discriminative metric to distinguish between clean and various types of poisoned samples (e.g., various triggers, various poisoning ratios). Inspired by a common phenomenon in backdoor attacks that the backdoored model tend to map significantly different poisoned and clean samples within the target class to similar activation areas, we introduce a novel perspective of the circular distribution of the gradients w.r.t. sample activation, dubbed gradient circular distribution (GCD). And, we find two interesting observations based on GCD. One is that the GCD of samples in the target class is much more dispersed than that in the clean class. The other is that in the GCD of target class, poisoned and clean samples are clearly separated. Inspired by above two observations, we develop an innovative three-stage poisoned sample detection approach, called Activation Gradient based Poisoned sample Detection (AGPD). First, we calculate GCDs of all classes from the model trained on the untrustworthy dataset. Then, we identify the target class(es) based on the difference on GCD dispersion between target and clean classes. Last, we filter out poisoned samples within the identified target class(es) based on the clear separation between poisoned and clean samples. Extensive experiments under various settings of backdoor attacks demonstrate the superior detection performance of the proposed method to existing poisoned detection approaches according to sample activation-based metrics.

We study streaming algorithms for proportionally fair clustering, a notion originally suggested by Chierichetti et al. (2017), in the sliding window model. We show that although there exist efficient streaming algorithms in the insertion-only model, surprisingly no algorithm can achieve finite ratio without violating the fairness constraint in sliding window. Hence, the problem of fair clustering is a rare separation between the insertion-only streaming model and the sliding window model. On the other hand, we show that if the fairness constraint is relaxed by a multiplicative $(1+\varepsilon)$ factor, there exists a $(1 + \varepsilon)$-approximate sliding window algorithm that uses $\text{poly}(k\varepsilon^{-1}\log n)$ space. This achieves essentially the best parameters (up to degree in the polynomial) provided the aforementioned lower bound. We also implement a number of empirical evaluations on real datasets to complement our theoretical results.

Lifelong learning algorithms enable models to incrementally acquire new knowledge without forgetting previously learned information. Contrarily, the field of machine unlearning focuses on explicitly forgetting certain previous knowledge from pretrained models when requested, in order to comply with data privacy regulations on the right-to-be-forgotten. Enabling efficient lifelong learning with the capability to selectively unlearn sensitive information from models presents a critical and largely unaddressed challenge with contradicting objectives. We address this problem from the perspective of simultaneously preventing catastrophic forgetting and allowing forward knowledge transfer during task-incremental learning, while ensuring exact task unlearning and minimizing memory requirements, based on a single neural network model to be adapted. Our proposed solution, privacy-aware lifelong learning (PALL), involves optimization of task-specific sparse subnetworks with parameter sharing within a single architecture. We additionally utilize an episodic memory rehearsal mechanism to facilitate exact unlearning without performance degradations. We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacy-aware unlearning mechanisms for responsible AI applications.

As machine learning models grow in complexity and increasingly rely on publicly sourced data, such as the human-annotated labels used in training large language models, they become more vulnerable to label poisoning attacks. These attacks, in which adversaries subtly alter the labels within a training dataset, can severely degrade model performance, posing significant risks in critical applications. In this paper, we propose $\textbf{Floral}$, a novel adversarial training defense strategy based on support vector machines (SVMs) to counter these threats. Utilizing a bilevel optimization framework, we cast the training process as a non-zero-sum Stackelberg game between an $\textit{attacker}$, who strategically poisons critical training labels, and the $\textit{model}$, which seeks to recover from such attacks. Our approach accommodates various model architectures and employs a projected gradient descent algorithm with kernel SVMs for adversarial training. We provide a theoretical analysis of our algorithm’s convergence properties and empirically evaluate $\textbf{Floral}$'s effectiveness across diverse classification tasks. Compared to robust baselines and foundation models such as RoBERTa, $\textbf{Floral}$ consistently achieves higher robust accuracy under increasing attacker budgets. These results underscore the potential of $\textbf{Floral}$ to enhance the resilience of machine learning models against label poisoning threats, thereby ensuring robust classification in adversarial settings.

The widespread deployment of pre-trained language models (PLMs) has exposed them to textual backdoor attacks, particularly those planted during the pre-training stage. These attacks pose significant risks to high-reliability applications, as they can stealthily affect multiple downstream tasks. While certifying robustness against such threats is crucial, existing defenses struggle with the high-dimensional, interdependent nature of textual data and the lack of access to original poisoned pre-training data. To address these challenges, we introduce **F**uzzed **R**andomized **S**moothing (**FRS**), a novel approach for efficiently certifying language model robustness against backdoor attacks. FRS integrates software robustness certification techniques with biphased model parameter smoothing, employing Monte Carlo tree search for proactive fuzzing to identify vulnerable textual segments within the Damerau-Levenshtein space. This allows for targeted and efficient text randomization, while eliminating the need for access to poisoned training data during model smoothing. Our theoretical analysis demonstrates that FRS achieves a broader certified robustness radius compared to existing methods. Extensive experiments across various datasets, model configurations, and attack strategies validate FRS's superiority in terms of defense efficiency, accuracy, and robustness.

Large Language Models (LLMs) have shown significant promise in real-world decision-making tasks for embodied artificial intelligence, especially when fine-tuned to leverage their inherent common sense and reasoning abilities while being tailored to specific applications. However, this fine-tuning process introduces considerable safety and security vulnerabilities, especially in safety-critical cyber-physical systems. In this work, we propose the first comprehensive framework for **B**ackdoor **A**ttacks against **L**LM-based **D**ecision-making systems (BALD) in embodied AI, systematically exploring the attack surfaces and trigger mechanisms. Specifically, we propose three distinct attack mechanisms: *word injection*, *scenario manipulation*, and *knowledge injection*, targeting various components in the LLM-based decision-making pipeline. We perform extensive experiments on representative LLMs (GPT-3.5, LLaMA2, PaLM2) in autonomous driving and home robot tasks, demonstrating the effectiveness and stealthiness of our backdoor triggers across various attack channels, with cases like vehicles accelerating toward obstacles and robots placing knives on beds. Our word and knowledge injection attacks achieve nearly 100\% success rate across multiple models and datasets while requiring only limited access to the system. Our scenario manipulation attack yields success rates exceeding 65\%, reaching up to 90\%, and does not require any runtime system intrusion. We also assess the robustness of these attacks against defenses, revealing their resilience. Our findings highlight critical security vulnerabilities in embodied LLM systems and emphasize the urgent need for safeguarding these systems to mitigate potential risks.

Large language models (LLMs) have gained human trust due to their capabilities and helpfulness. However, this in turn may allow LLMs to affect users' mindsets by manipulating language. It is termed as gaslighting, a psychological effect. In this work, we aim to investigate the vulnerability of LLMs under prompt-based and fine-tuning-based gaslighting attacks. Therefore, we propose a two-stage framework DeepCoG designed to: 1) elicit gaslighting plans from LLMs with the proposed DeepGaslighting prompting template, and 2) acquire gaslighting conversations from LLMs through our Chain-of-Gaslighting method. The gaslighting conversation dataset along with a corresponding safe dataset is applied to fine-tuning-based attacks on open-source LLMs and anti-gaslighting safety alignment on these LLMs. Experiments demonstrate that both prompt-based and fine-tuning-based attacks transform three open-source LLMs into gaslighters. In contrast, we advanced three safety alignment strategies to strengthen~(by $12.05\%$) the safety guardrail of LLMs. Our safety alignment strategies have minimal impacts on the utility of LLMs. Empirical studies indicate that an LLM may be a potential gaslighter, even if it passed the harmfulness test on general dangerous queries.

We study a setting of collecting and learning from private data distributed across end users. In the shuffled model of differential privacy, the end users partially protect their data locally before sharing it, and their data is also anonymized during its collection to enhance privacy. This model has recently become a prominent alternative to central DP, which requires full trust in a central data curator, and local DP, where fully local data protection takes a steep toll on downstream accuracy. Our main technical result is a shuffled DP protocol for privately estimating the kernel density function of a distributed dataset, with accuracy essentially matching central DP. We use it to privately learn a classifier from the end user data, by learning a private density function per class. Moreover, we show that the density function itself can recover the semantic content of its class, despite having been learned in the absence of any unprotected data. Our experiments show the favorable downstream performance of our approach, and highlight key downstream considerations and trade-offs in a practical ML deployment of shuffled DP.

The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a humanized proxy-attack (HUMPA) strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8x7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 95.0% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model.

Diffusion models have a tendency to exactly replicate their training data, especially when trained on small datasets. Most prior work has sought to mitigate this problem by imposing differential privacy constraints or masking parts of the training data, resulting in a notable substantial decrease in image quality. We present CPSample, a method that modifies the sampling process to prevent training data replication while preserving image quality. CPSample utilizes a classifier that is trained to overfit on random binary labels attached to the training data. CPSample then uses classifier guidance to steer the generation process away from the set of points that can be classified with high certainty, a set that includes the training data. CPSample achieves FID scores of 4.97 and 2.97 on CIFAR-10 and CelebA-64, respectively, without producing exact replicates of the training data. Unlike prior methods intended to guard the training images, CPSample only requires training a classifier rather than retraining a diffusion model, which is computationally cheaper. Moreover, our technique provides diffusion models with greater robustness against membership inference attacks, wherein an adversary attempts to discern which images were in the model's training dataset. We show that CPSample behaves like a built-in rejection sampler, and we demonstrate its capabilities to prevent mode collapse in Stable Diffusion.

Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within adversarial prompts. While most existing defenses attempt to mitigate the effects of adversarial prompts, they often prove inadequate as adversarial prompts can take arbitrary, adaptive forms. This paper introduces RobustKV, a novel jailbreak defense that takes a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for an adversarial prompt to be effective, its tokens must achieve sufficient `importance' (measured by attention scores), which consequently lowers the importance of tokens in the concealed harmful query. Therefore, by carefully evicting the KVs of low-ranked tokens, RobustKV minimizes the harmful query's presence in the KV cache, thus preventing the LLM from generating informative responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's performance on benign queries. Notably, RobustKV creates an interesting effectiveness-evasiveness dilemma for the adversary, leading to its robustness against adaptive attacks.{(Warning: This paper contains potentially harmful content generated by LLMs.)}