Lune

ACL Findings2024

BadActs: A Universal Backdoor Defense in the Activation Space

Biao Yi, Sishuo Chen, Yiming Li, Tong Li, Baolei Zhang, Zheli Liu

Abstract

Backdoor attacks pose an increasingly severe security threat to Deep Neural Networks (DNNs) during their development stage. In response, backdoor sample purification has emerged as a promising defense mechanism, aiming to eliminate backdoor triggers while preserving the integrity of the clean content in the samples. However, existing approaches have been predominantly focused on the word space, which are ineffective against featurespace triggers and significantly impair performance on clean data. To address this, we introduce a universal backdoor defense that purifies backdoor samples in the activation space by drawing abnormal activations towards optimized minimum clean activation distribution intervals. The advantages of our approach are twofold: (1) By operating in the activation space, our method captures from surface-level information like words to higher-level semantic concepts such as syntax, thus counteracting diverse triggers; (2) the fine-grained continuous nature of the activation space allows for more precise preservation of clean content while removing triggers. Furthermore, we propose a detection module based on statistical information of abnormal activations, to achieve a better trade-off between clean accuracy and defending performance. Extensive experiments on diverse datasets and against diverse attacks (including syntax and style attacks) demonstrate that our defense achieves state-of-the-art performance. 1

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

11 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning cites Weight Poisoning Attacks on Pretrained Models
  • Locating and Editing Factual Associations in GPT cites Language Models are Few-Shot Learners
  • DUP: Detection-guided Unlearning for Backdoor Purification in Language Models cites Weight Poisoning Attacks on Pretrained Models
  • Defending against Backdoor Attacks via Module Switching cites Weight Poisoning Attacks on Pretrained Models
  • Poisoning Language Models During Instruction Tuning cites Language Models are Few-Shot Learners
  • Poisoning Language Models During Instruction Tuning cites Weight Poisoning Attacks on Pretrained Models
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers11

  1. DUP: Detection-guided Unlearning for Backdoor Purification in Language ModelsMan Hu, Yahui Ding, Yatao Yang, Liangyu Chen et al.AAAI 2026 · 4 citations
  2. Defending against Backdoor Attacks via Module SwitchingWeijun Li, Ansh Arora, Xuanli He, Mark Dras et al.ICLR 2026 · 2 citations
  3. Efficient Input-Level Backdoor Defense on Text-to-Image Synthesis via Neuron Activation VariationShengfang Zhai, Jiajun Li, Yue Liu, Huanran Chen et al.ICCV 2025 · 2 citations
  4. RepGuard: Adaptive Feature Decoupling for Robust Backdoor Defense in Large Language ModelsChenxu Niu, Jie M. Zhang, Yanbing Liu, Yunpeng Li et al.NeurIPS 2025 · 1 citation
  5. Don't Shift the Trigger: Robust Gradient Ascent for Backdoor UnlearningXingyi Zhao, Tian Xie, Xiaojun Qi, Depeng Xu et al.ICLR 2026

Builds on29

  1. Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah et al.NeurIPS 2020 · 64,255 citations
  2. Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
  3. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than GeneratorsKevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. ManningICLR 2020 · 541 citations
  4. Poisoning Language Models During Instruction TuningAlexander Wan, Eric Wallace, Sheng Shen, Dan KleinICML 2023 · 319 citations
  5. Weight Poisoning Attacks on Pretrained ModelsKeita Kurita, Paul Michel, Graham NeubigACL 2020 · 312 citations

Related papers

  1. Purifier: Plug-and-play Backdoor Mitigation for Pre-trained Models Via Anomaly Activation SuppressionXiaoyu Zhang, Yulin Jin, Tao Wang, Jian Lou et al.ACM MM 2022 · 12 citations
  2. Prototype Guided Backdoor Defense via Activation Space ManipulationVenkat Adithya Amula, Sunayana Samavedam, Saurabh Saini, Avani Gupta et al.ICCV 2025 · 2 citations
  3. Redeem Myself: Purifying Backdoors in Deep Learning Models using Self Attention DistillationXueluan Gong, Yanjiao Chen, Wang Yang, Qian Wang et al.S&P 2023
  4. Beating Backdoor Attack at Its Own GameMin Liu, Alberto L. Sangiovanni-Vincentelli, Xiangyu YueICCV 2023 · 19 citations
  5. Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor DefenseRui Min, Zeyu Qin, Nevin L. Zhang, Li Shen et al.NeurIPS 2024 · 16 citations
Dusk over the sea between two cliffs drawn in fine vertical lines