LLMs demonstrate remarkable utility but remain vulnerable to jailbreak attacks that aim to elicit harmful responses. Existing defenses, including post-training alignment and prompt engineering, rely on training on safety-annotated datasets and safe prompt templates, struggling with adaptability to out-of-distribution (OOD) attacks. Steering internal representations of LLMs provides real-time adjustments to defend against OOD attacks. However, it struggles with maintaining model utility, since modifying the representation disrupts the forward pass of inference. It barely considers the competitive objectives of helpfulness and harmlessness in LLMs. We argue that adversarial game-based approaches promise a solution for conflicts between the two objectives. In this paper, we propose **A**dversarial **G**ame **D**efense (AGD), an adversarial game-based defense method that dynamically adjusts LLMs’ internal representations to achieve a balanced trade-off between helpfulness and harmlessness. AGD first proposes an interquartile range (IQR) method to detect abnormal attention weights and correct the abnormal weights via adversarial training. AGD adopts a bi-level optimization to play a two-player variable-sum game to approach Nash Equilibrium (NE), where the two players adversarially refine head activations for helpfulness and harmlessness respectively. Furthermore, AGD applies an expert model to next-token sampling to generate safer responses. Experiments show that AGD significantly improves LLMs’ safety over all baselines.

Current backdoor attack defenders in Natural Language Processing (NLP) typically involve data reduction or model pruning, risking losing crucial information. To address this challenge, we introduce a novel backdoor defender, i.e., BadWindtunnel, in which we build a high-noise simulated training environment, similar to the wind tunnel, which allows precise control over training conditions to model the backdoor learning behavior without affecting the final model. We also use the confidence variance as a learning behavior quantification metric in the simulated training, which is based on the characteristics of backdoor-poisoned data (shorted in poisoned data): higher learnability and robustness. In addition, we propose a two-step strategy to further model poisoned data, including target label identification and poisoned data revealing. Extensive experiments demonstrate BadWindtunnel’s superiority, with a 21% higher average reduction in attack success rate than the second-best defender.

Using large language models (LLMs) has a potential risk of privacy leakage since the data with sensitive information may be used for fine-tuning the LLMs. Differential privacy (DP) provides theoretical guarantees of privacy protection, but its practical application in LLMs still has the problem of privacy-utility trade-off. Researchers synthesized data with strong generation capabilities closed-source LLMs (i.e., GPT-4) under DP to alleviate this problem, but this method is not so flexible in fitting the given privacy distributions without fine-tuning. Besides, such methods can hardly balance the diversity of synthetic data and its relevance to target privacy data without accessing so much private data. To this end, this paper proposes DPGA-TextSyn, combining general LLMs with genetic algorithm (GA) to produce relevant and diverse synthetic text under DP constraints. First, we integrate the privacy gene (i.e., metadata) to generate better initial samples. Then, to achieve survival of the fittest and avoid homogeneity, we use privacy nearest neighbor voting and similarity suppression to select elite samples. In addition, we expand elite samples via genetic strategies such as mutation, crossover, and generation to expand the search scope of GA. Experiments show that this method significantly improves the performance of the model in downstream tasks while ensuring privacy.