Nay Myat Min

Nay Myat Min, Long H. Pham, Yige Li et al.

Large language models (LLMs) demonstrate remarkable performance across myriad language tasks, yet they remain vulnerable to backdoor attacks, where adversaries implant hidden triggers that systematically manipulate model outputs. Traditional defenses focus on explicit token-level anomalies and therefore overlook semantic backdoors-covert triggers embedded at the conceptual level (e.g., ideological stances or cultural references) that rely on meaning-based cues rather than lexical oddities. We first show, in a controlled finetuning setting, that such semantic backdoors can be implanted with only a small poisoned corpus, establishing their practical feasibility. We then formalize the notion of semantic backdoors in LLMs and introduce a black-box detection framework, RAVEN (short for "Response Anomaly Vigilance for uncovering semantic backdoors"), which combines semantic entropy with cross-model consistency analysis. The framework probes multiple models with structured topic-perspective prompts, clusters the sampled responses via bidirectional entailment, and flags anomalously uniform outputs; cross-model comparison isolates model-specific anomalies from corpus-wide biases. Empirical evaluations across diverse LLM families (GPT-4o, Llama, DeepSeek, Mistral) uncover previously undetected semantic backdoors, providing the first proof-of-concept evidence of these hidden vulnerabilities and underscoring the urgent need for concept-level auditing of deployed language models. We open-source our code and data at https://github.com/NayMyatMin/RAVEN.

Nay Myat Min, Long H. Pham, Jun Sun

Deep neural networks have achieved remarkable success across various applications; however, their vulnerability to backdoor attacks poses severe security risks -- especially in situations where only a limited set of clean samples is available for defense. In this work, we address this critical challenge by proposing ULRL (UnLearn and ReLearn for backdoor removal), a novel two-phase approach for comprehensive backdoor removal. Our method first employs an unlearning phase, in which the network's loss is intentionally maximized on a small clean dataset to expose neurons that are excessively sensitive to backdoor triggers. Subsequently, in the relearning phase, these suspicious neurons are recalibrated using targeted reinitialization and cosine similarity regularization, effectively neutralizing backdoor influences while preserving the model's performance on benign data. Extensive experiments with 12 backdoor types on multiple datasets (CIFAR-10, CIFAR-100, GTSRB, and Tiny-ImageNet) and architectures (PreAct-ResNet18, VGG19-BN, and ViT-B-16) demonstrate that ULRL significantly reduces the attack success rate without compromising clean accuracy -- even when only 1% of clean data is used for defense.