后门攻击论文趋势 2023–2026

主要结论

• 后门攻击论文从 2023 年的 73 篇增长到 2026 年的 111 篇，三年间扩大到约 1.5 倍。
• NeurIPS 是发后门攻击论文最多的会议，2023–2026 年共 124 篇，占四个会议总量的 33%。
• 交叉方向里涨得最快的是对齐与安全：同时属于两个主题的论文从 2023 年的 0 篇增至 2026 年的 16 篇。
• 在「安全与可信」的 5 个主题里，它的 2026 年论文数排第 2；同组增长最快的水印是 9.3 倍，后门攻击是 1.5 倍。

代表论文

按年份和会议均匀挑选，方便快速看到每一届的样貌

• NeurIPS 2026A Theoretical Analysis of Backdoor Learning as Simplicity-Biased Optimization Dynamics
• ICML 2026Coupled Trigger Optimization and Vulnerable Parameter Alignment for Persistent Backdoor Attacks on Federated Learning
• ICLR 2026Defending against Backdoor Attacks via Module Switching
• CVPR 2026Eliminate Distance Differences Induced by Backdoor Attacks Layer-Selective Training and
• NeurIPS 2025A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers
• ICML 2025A Closer Look at Backdoor Attacks on CLIP
• ICLR 2025Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks
• CVPR 2025BadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models
• NeurIPS 2024Reproducibility Study on Adversarial Attacks Against Robust Transformer Trackers
• ICML 2024Better Safe than Sorry: Pre-training CLIP against Targeted Data Poisoning and Backdoor Attacks
• ICLR 2024Adversarial Feature Map Pruning for Backdoor
• CVPR 2024Data Poisoning based Backdoor Attacks to Contrastive Learning
• NeurIPS 2023A Unified Detection Framework for Inference-Stage Backdoor Defenses
• ICML 2023Data Poisoning Attacks Against Multimodal Encoders
• ICLR 2023Adversarial Attacks on Adversarial Bandits
• CVPR 2023Architectural Backdoors in Neural Networks
• NeurIPS 2026AI Models Can Provably Hide Arbitrary Capabilities
• ICML 2026DF-LoGiT: Data-Free Logic-Gated Backdoor Attacks in Vision Transformers
• ICLR 2026Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning
• CVPR 2026Logit-Margin Repulsion for Backdoor Defense

全部后门攻击论文

共 110 篇，第 2 / 3 页

51. ICML 2026From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs
52. ICML 2026Improving the Sensitivity of Backdoor Detectors via Class Subspace Orthogonalization
53. ICML 2026INDEXGUARD: Index-only Backdoor Vetting for Secure Federated PEFT of Large Language Models
54. ICML 2026Neural Honeytrace: Plug&Play Watermarking Framework against Model Extraction Attacks
55. ICML 2026New Wide-Net-Casting Jailbreak Attacks Risk Large Models
56. ICML 2026PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks
57. ICML 2026Persistent Backdoor Attacks in Class-Incremental Learning via Structural Invariant Anchoring
58. ICML 2026PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent
59. ICML 2026Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework
60. ICML 2026Rethinking Forgery Attacks on Semantic Watermarks in Black-Box Settings: A Geometric Distortion Perspective
61. ICML 2026Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models
62. ICML 2026Robust In-Context Reinforcement Learning Under Reward Poisoning Attacks
63. ICML 2026Semantic-level Backdoor Attack against Text-to-Image Diffusion Models
64. ICML 2026SemBind: Binding Diffusion Watermarks to Semantics Against Black-Box Forgery Attacks
65. ICML 2026SkillTrojan: Backdoor Attacks on Skill-Based Agent Systems
66. ICML 2026SORA: Free Second-Order Attacks in Fast Adversarial Training
67. ICML 2026TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning
68. ICML 2026Theory of Continual Learning Against Data Poisoning Attacks
69. ICML 2026Theory of Minimal Weight Perturbations in Deep Networks and its Applications for Low-Rank Activated Backdoor Attacks
70. ICML 2026TimeGuard: Channel-wise Pool Training for Backdoor Defense in Time Series Forecasting
71. ICML 2026TokenSwap: Backdoor Attack on the Compositional Understanding of Large Vision-Language Models
72. ICML 2026Toward Subspace-Perturbed Trajectory-Aware Backdoor Attacks in Deep Reinforcement Learning
73. ICML 2026Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges
74. ICML 2026Uncovering Hidden Triggers: Backdoor Attribution in Language Models
75. NeurIPS 2026A Theoretical Analysis of Backdoor Learning as Simplicity-Biased Optimization Dynamics
76. NeurIPS 2026AI Models Can Provably Hide Arbitrary Capabilities
77. NeurIPS 2026Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation
78. NeurIPS 2026Backdoor Channels Hidden in Latent Space: Cryptographic Undetectability in Modern Neural Networks
79. NeurIPS 2026Backdoor Purification for LoRA-Tuned LLMs via Null-Space Projection
80. NeurIPS 2026Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models
81. NeurIPS 2026CASPIAN: Online Detection and Attribution of Cascade Attacks in LLM Multi-Agent Systems via Cross-Channel Causal Monitoring
82. NeurIPS 2026Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse
83. NeurIPS 2026Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers
84. NeurIPS 2026Consistency-Verified Backdoor Defense for Federated Graph Learning via Cross-Layer Drift
85. NeurIPS 2026DetectViT: Test-time Backdoor Detection for Vision Transformers via Inter-Head Attention Discrepancy
86. NeurIPS 2026Evaluating an Evaluation: Membership Inference Attacks as Machine Unlearning Diagnostics
87. NeurIPS 2026ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
88. NeurIPS 2026FloatDoor: Platform triggered Backdoors in LLMs
89. NeurIPS 2026HPE: Hallucinated Positive Entanglement for Backdoor Attacks in Federated Self-Supervised Learning
90. NeurIPS 2026Latent-space Attacks for Refusal Evasion in Language Models
91. NeurIPS 2026Learning Robust Representations for Defending White-Box Adversarial Attacks in Continual Learning
92. NeurIPS 2026Leveraging Soft Prompts for Privacy Attacks in Federated Prompt Tuning
93. NeurIPS 2026Local FDR Membership Inference Attacks: Multiple Testing and the Role of Ridge Regularization
94. NeurIPS 2026Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing
95. NeurIPS 2026Not Suppressing or Purifying: Backdoor Containment via Expert Quarantine and Shutdown in LLMs
96. NeurIPS 2026Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
97. NeurIPS 2026Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples
98. NeurIPS 2026Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection
99. NeurIPS 2026Provable Robustness against Backdoor Attacks via the Primal-Dual Perspective on Differential Privacy
100. NeurIPS 2026Robust and Efficient Backdoor Mitigation for ML Models via Tolerant Property Testing

常见问题

2023–2026 年四大会议一共有多少篇后门攻击论文？

共 381 篇，按年份分别是 2023 年 73 篇、2024 年 105 篇、2025 年 92 篇、2026 年 111 篇。

NeurIPS 2026 有多少篇后门攻击论文？

NeurIPS 2026 有 36 篇后门攻击论文，上一届是 31 篇。

ICLR 2026 有多少篇后门攻击论文？

ICLR 2026 有 20 篇后门攻击论文，上一届是 25 篇。

ICML 2026 有多少篇后门攻击论文？

ICML 2026 有 36 篇后门攻击论文，上一届是 17 篇。

CVPR 2026 有多少篇后门攻击论文？

CVPR 2026 有 19 篇后门攻击论文，上一届是 19 篇。

哪个会议的后门攻击论文最多？

NeurIPS，2023–2026 年共 124 篇。