Research map / Adversarial robustness and security / Backdoor and poisoning
Qu-ANTI-zation: Exploiting Quantization Artifacts for Achieving Adversarial Outcomes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add quantization-triggered indiscriminate attack objective | ResNet18 · CIFAR10, 8-bit | accuracy: 93.6% -> 10.0% | Hurt |
| Add quantization-triggered indiscriminate attack objective | VGG16 · Tiny ImageNet, 8-bit | accuracy: 42.9% -> 0.6% | Hurt |
| Restrict attack loss to target class | VGG16 · CIFAR10 target class, 8-bit | accuracy: 93.5% -> 9.4% | Hurt |
| Targeted misclassification attack on a specific sample | ResNet18 · CIFAR10, 4-bit | target-label success: 20.0% -> 100% | Helped |
| Quantization-activated backdoor vs standard backdoor | ResNet18 · CIFAR10, 32-bit | backdoor success rate: 98.3% -> 11.3% | Helped |
| Quantize backdoored model to 4-bit | ResNet18 · CIFAR10 | backdoor success rate: 11.3% -> 100% | Helped |
| Victim uses OCS robust quantization | Compromised ResNet18 · CIFAR10, 8-bit | backdoor success rate: 12.8% -> 99.7% | No clear effect |
| Victim uses OMSE; disable activation clipping | Compromised ResNet18 · CIFAR10, 8-bit | backdoor success rate: 25.0% -> 88% | Mixed |
| Victim fine-tunes compromised model | Compromised ResNet18 · CIFAR10, 8-bit | accuracy: 93.4% | Hurt |
| Victim adds random noise to parameters | Compromised ResNet18 · CIFAR10, 4-bit | accuracy: 13.1% | No clear effect |
| Transfer learning: re-train only last layer of compromised teacher | ResNet18 student · CIFAR10 from Tiny ImageNet, 4-bit | accuracy: ~74% -> 20--24% | No clear effect |
| Federated learning with 5 malicious of 100 participants | AlexNet · CIFAR10, 4-bit | backdoor success rate: 20 to 81\% | Helped |
Closest papers by meaning
- Mind the Gap: A Practical Attack on GGUF Quantization (ICML 2025 · similarity 0.88)
- Purifying Quantization-conditioned Backdoors via Layer-wise Activation Correction with Distribution Approximation (ICML 2024 · similarity 0.88)
- Nearest is Not Dearest: Towards Practical Defense against Quantization-conditioned Backdoor Attacks (CVPR 2024 · similarity 0.88)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.87)
- Understanding the Unfairness in Network Quantization (ICML 2025 · similarity 0.85)
- One-Step Forward and Backtrack: Overcoming Zig-Zagging in Loss-Aware Quantization Training (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.