Large Language Models (LLMs) commonly rely on explicit refusal prefixes for safety, making them vulnerable to prefix injection attacks. We introduce HumorReject, a novel data-driven approach that fundamentally reimagines LLM safety by decoupling it from refusal prefixes through the use of humor as an indirect refusal strategy. Rather than explicitly rejecting harmful instructions, HumorReject responds with contextually appropriate humor that naturally defuses potentially dangerous requests while maintaining engaging interactions. Our approach effectively addresses the common "over-defense" issues in existing safety mechanisms, demonstrating superior robustness against various attack vectors while preserving natural and high-quality interactions on legitimate tasks. Our findings suggest that innovations at the data level are even more fundamental than the alignment algorithm itself in achieving effective LLM safety, opening new directions for developing more resilient and user-friendly AI systems. Our code and dataset are available at https://github.com/wooozihui/HumorReject
Adversarial examples (AEs) are an additive amalgamation of clean examples and artificially malicious perturbations. Attackers often leverage random noise and multiple random restarts to initialize perturbation starting points, thereby increasing the diversity of AEs. Given the non-convex nature of the loss function, employing randomness to augment the attack's success rate may lead to considerable computational overhead. To overcome this challenge, we introduce the one-hot mean square error loss to guide the initialization. This loss is combined with the strongest first-order attack, the projected gradient descent, alongside a dynamic attack step size adjustment strategy to form a comprehensive attack process. Through experimental validation, we demonstrate that our method outperforms baseline attacks in constrained attack budget scenarios and regular experimental settings. This establishes it as a reliable measure for assessing the robustness of deep learning models. We explore the broader application of this initialization strategy in enhancing the defense impact of few-shot classification models. We aspire to provide valuable insights for the community in designing attack and defense mechanisms.
In this paper, we investigate on improving the adversarial robustness obtained in adversarial training (AT) via reducing the difficulty of optimization. To better study this problem, we build a novel Bregman divergence perspective for AT, in which AT can be viewed as the sliding process of the training data points on the negative entropy curve. Based on this perspective, we analyze the learning objectives of two typical AT methods, i.e., PGD-AT and TRADES, and we find that the optimization process of TRADES is easier than PGD-AT for that TRADES separates PGD-AT. In addition, we discuss the function of entropy in TRADES, and we find that models with high entropy can be better robustness learners. Inspired by the above findings, we propose two methods, i.e., FAIT and MER, which can both not only reduce the difficulty of optimization under the 10-step PGD adversaries, but also provide better robustness. Our work suggests that reducing the difficulty of optimization under the 10-step PGD adversaries is a promising approach for enhancing the adversarial robustness in AT.
\emph{Consistent teaching} is an effective paradigm for implementing knowledge distillation (KD), where both student and teacher models receive identical inputs, and KD is treated as a function matching task (FunMatch). However, one limitation of FunMatch is that it does not account for the transfer of adversarial robustness, a model's resistance to adversarial attacks. To tackle this problem, we propose a simple but effective strategy called Adversarial Function Matching (AdvFunMatch), which aims to match distributions for all data points within the $\ell_p$-norm ball of the training data, in accordance with consistent teaching. Formulated as a min-max optimization problem, AdvFunMatch identifies the worst-case instances that maximizes the KL-divergence between teacher and student model outputs, which we refer to as "mismatched examples," and then matches the outputs on these mismatched examples. Our experimental results show that AdvFunMatch effectively produces student models with both high clean accuracy and robustness. Furthermore, we reveal that strong data augmentations (\emph{e.g.}, AutoAugment) are beneficial in AdvFunMatch, whereas prior works have found them less effective in adversarial training. Code is available at \url{https://gitee.com/zihui998/adv-fun-match}.