Yiming Li

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.

In recent years, RAG has emerged as a key paradigm for enhancing large language models (LLMs). By integrating externally retrieved information, RAG alleviates issues like outdated knowledge and, crucially, insufficient domain expertise. While effective, RAG introduces new risks of external data extraction attacks (EDEAs), where sensitive or copyrighted data in its knowledge base may be extracted verbatim. These risks are particularly acute when RAG is used to customize specialized LLM applications with private knowledge bases. Despite initial studies exploring these risks, they often lack a formalized framework, robust attack performance, and comprehensive evaluation, leaving critical questions about real-world EDEA feasibility unanswered. In this paper, we present the first comprehensive study to formalize EDEAs against retrieval-augmented LLMs. We first formally define EDEAs and propose a unified framework decomposing their design into three components: extraction instruction, jailbreak operator, and retrieval trigger, under which prior attacks can be considered instances within our framework. Guided by this framework, we develop SECRET: a Scalable and EffeCtive exteRnal data Extraction aTtack. Specifically, SECRET incorporates (1) an adaptive optimization process using LLMs as optimizers to generate specialized jailbreak prompts for EDEAs, and (2) cluster-focused triggering, an adaptive strategy that alternates between global exploration and local exploitation to efficiently generate effective retrieval triggers. Extensive evaluations across 4 models reveal that SECRET significantly outperforms previous attacks, and is highly effective against all 16 tested RAG instances. Notably, SECRET successfully extracts 35% of the data from RAG powered by Claude 3.7 Sonnet for the first time, whereas other attacks yield 0% extraction. Our findings call for attention to this emerging threat.

Large-scale vision-language models, especially CLIP, have demonstrated remarkable performance across diverse downstream tasks. Soft prompts, as carefully crafted modules that efficiently adapt vision-language models to specific tasks, necessitate effective copyright protection. In this paper, we investigate model copyright protection by auditing whether suspicious third-party models incorporate protected soft prompts. While this can be viewed as a special case of model ownership auditing, our analysis shows that existing techniques are ineffective due to prompt learning's unique characteristics. Non-intrusive auditing is inherently prone to false positives when independent models share similar data distributions with victim models. Intrusive approaches also fail: backdoor methods designed for CLIP cannot embed functional triggers, while extending traditional DNN backdoor techniques to prompt learning suffers from harmfulness and ambiguity challenges. We find that these failures in intrusive auditing stem from the same fundamental reason: watermarking operates within the same decision space as the primary task yet pursues opposing objectives. Motivated by these findings, we propose sequential watermarking for soft prompts (SWAP), which implants watermarks into a different and more complex space. SWAP encodes watermarks through a specific order of defender-specified out-of-distribution classes, inspired by the zero-shot prediction capability of CLIP. This watermark, which is embedded in a more complex space, keeps the original prediction label unchanged, making it less opposed to the primary task. We further design a hypothesis-test-guided verification protocol for SWAP and provide theoretical analyses of success conditions. Extensive experiments on 11 datasets demonstrate SWAP's effectiveness, harmlessness, and robustness against potential adaptive attacks.