Yiming Li

Yukun Chen, Boheng Li, Yu Yuan et al.

Knowledge distillation (KD) is a vital technique for deploying deep neural networks (DNNs) on resource-constrained devices by transferring knowledge from large teacher models to lightweight student models. While teacher models from third-party platforms may undergo security verification (\eg, backdoor detection), we uncover a novel and critical threat: distillation-conditional backdoor attacks (DCBAs). DCBA injects dormant and undetectable backdoors into teacher models, which become activated in student models via the KD process, even with clean distillation datasets. While the direct extension of existing methods is ineffective for DCBA, we implement this attack by formulating it as a bilevel optimization problem and proposing a simple yet effective method (\ie, SCAR). Specifically, the inner optimization simulates the KD process by optimizing a surrogate student model, while the outer optimization leverages outputs from this surrogate to optimize the teacher model for implanting the conditional backdoor. Our SCAR addresses this complex optimization utilizing an implicit differentiation algorithm with a pre-optimized trigger injection function. Extensive experiments across diverse datasets, model architectures, and KD techniques validate the effectiveness of our SCAR and its resistance against existing backdoor detection, highlighting a significant yet previously overlooked vulnerability in the KD process. Our code is available at https://github.com/WhitolfChen/SCAR.

Wenjie Li, Siying Gu, Yiming Li et al.

Backdoor detection is currently the mainstream defense against backdoor attacks in federated learning (FL), where malicious clients upload poisoned updates that compromise the global model and undermine the reliability of FL deployments. Existing backdoor detection techniques fall into two categories, including passive and proactive ones, depending on whether the server proactively modifies the global model. However, both have inherent limitations in practice: passive defenses are vulnerable to common non-i.i.d. data distributions and random participation of FL clients, whereas current proactive defenses suffer inevitable out-of-distribution (OOD) bias because they rely on backdoor co-existence effects. To address these issues, we introduce a new proactive defense, dubbed Coward, inspired by our discovery of multi-backdoor collision effects, in which consecutively planted, distinct backdoors significantly suppress earlier ones. In general, we detect attackers by evaluating whether the server-injected, conflicting global watermark is erased during local training rather than retained. Our method preserves the advantages of proactive defenses in handling data heterogeneity (\ie, non-i.i.d. data) while mitigating the adverse impact of OOD bias through a revised detection mechanism. Extensive experiments on benchmark datasets confirm the effectiveness of Coward and its resilience to potential adaptive attacks. The code for our method would be available at https://github.com/still2009/cowardFL.

Xuankun Rong, Wenke Huang, Jian Liang et al.

Multimodal Large Language Models (MLLMs) are increasingly deployed in fine-tuning-as-a-service (FTaaS) settings, where user-submitted datasets adapt general-purpose models to downstream tasks. This flexibility, however, introduces serious security risks, as malicious fine-tuning can implant backdoors into MLLMs with minimal effort. In this paper, we observe that backdoor triggers systematically disrupt cross-modal processing by causing abnormal attention concentration on non-semantic regions--a phenomenon we term attention collapse. Based on this insight, we propose Believe Your Eyes (BYE), a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples. BYE operates via a three-stage pipeline: (1) extracting attention maps using the fine-tuned model, (2) computing entropy scores and profiling sensitive layers via bimodal separation, and (3) performing unsupervised clustering to remove suspicious samples. Unlike prior defenses, BYE equires no clean supervision, auxiliary labels, or model modifications. Extensive experiments across various datasets, models, and diverse trigger types validate BYE's effectiveness: it achieves near-zero attack success rates while maintaining clean-task performance, offering a robust and generalizable solution against backdoor threats in MLLMs.

Boheng Li, Renjie Gu, Junjie Wang et al.

Text-to-image (T2I) diffusion models have achieved impressive image generation quality and are increasingly fine-tuned for personalized applications. However, these models often inherit unsafe behaviors from toxic pretraining data, raising growing safety concerns. While recent safety-driven unlearning methods have made promising progress in suppressing model toxicity, they are identified to be fragile to downstream fine-tuning, where we reveal that state-of-the-art methods largely fail to retain their effectiveness even when fine-tuned on entirely benign datasets. To mitigate this problem, in this paper, we propose ResAlign, a safety-driven unlearning framework with enhanced resilience against downstream fine-tuning. By modeling downstream fine-tuning as an implicit optimization problem with a Moreau Envelope-based reformulation, ResAlign enables efficient gradient estimation to minimize the recovery of harmful behaviors. Additionally, a meta-learning strategy is proposed to simulate a diverse distribution of fine-tuning scenarios to improve generalization. Extensive experiments across a wide range of datasets, fine-tuning methods, and configurations demonstrate that ResAlign consistently outperforms prior unlearning approaches in retaining safety after downstream fine-tuning while preserving benign generation capability well.

Xiaobei Yan, Yiming Li, Hao Wang et al.

Large language models (LLMs) are widely deployed, but their growing compute demands expose them to inference cost attacks that maximize output length. We reveal that prior attacks are fundamentally self-targeting because they rely on crafted inputs, so the added cost accrues to the attacker's own queries and scales poorly in practice. In this work, we introduce the first bit-flip inference cost attack that directly modifies model weights to induce persistent overhead for all users of a compromised LLM. Such attacks are stealthy yet realistic in practice: for instance, in shared MLaaS environments, co-located tenants can exploit hardware-level faults (e.g., Rowhammer) to flip memory bits storing model parameters. We instantiate this attack paradigm with BitHydra, which (1) minimizes a loss that suppresses the end-of-sequence token (i.e., EOS) and (2) employs an efficient yet effective critical-bit search focused on the EOS embedding vector, sharply reducing the search space while preserving benign-looking outputs. We evaluate across 11 LLMs (1.5B-14B) under int8 and float16, demonstrating that our method efficiently achieves scalable cost inflation with only a few bit flips, while remaining effective even against potential defenses.