Research map / Adversarial robustness and security / Backdoor and poisoning
DISTIL: Data-Free Inversion of Suspicious Trojan Inputs via Latent Diffusion
Hossein Mirzaei, Zeinab Taghavi, Sepehr Rezaee, Masoud Hadi, Moein Madadi, Mackenzie W. Mathis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| search in diffusion latent space instead of pixel space (ablation) | DISTIL · TrojAI Round 4 | scanning accuracy: 60.5 -> 84.6 | Improved |
| search in diffusion latent space instead of pixel space (ablation) | DISTIL · TrojAI Round 11 | scanning accuracy: 57.4 -> 80.4 | Improved |
| add uniform noise injection at each reverse step (ablation) | DISTIL · TrojAI Round 1 | scanning accuracy: 80.5 -> 82.9 | Improved |
| add uniform noise injection (ablation) | DISTIL · TrojAI Round 4 | scanning accuracy: 81.6 -> 84.6 | Improved |
| use training-based hyperparameter selection instead of fixed lambda values (ablation) | DISTIL · TrojAI Round 2 | scanning accuracy: 72.8 -> 79.5 | Improved |
| use fast class pairing to cut cost from O(K^2) to O(K) (ablation) | DISTIL · TrojAI Round 0 | scanning accuracy: 83.1 -> 78.0 | Hurt |
| use fast class pairing (Round 3) (ablation) | DISTIL · TrojAI Round 3 | scanning accuracy: 78.4 -> 75.9 | Mixed |
| use ImageNet-pretrained diffusion backbone instead of GLIDE (ablation) | DISTIL · TrojAI Round 2 | scanning accuracy: 79.5 -> 78.4 | Hurt |
| use LSUN score-based diffusion backbone instead of GLIDE (ablation) | DISTIL · TrojAI Round 1 | scanning accuracy: 82.9 -> 74.2 | Hurt |
| add clean data supervision (hybrid conditioning) (ablation) | DISTIL · TrojAI Round 2 | scanning accuracy: 79.5 -> 82.4 | Improved |
| add clean data supervision (hybrid conditioning) (ablation) | DISTIL · TrojAI Round 11 | scanning accuracy: 80.4 -> 83.9 | Improved |
| DISTIL-Fast vs DISTIL runtime on TrojAI Round 11 (ablation) | DISTIL-Fast · TrojAI Round 11 | hours: 14.4 -> 0.2 | Improved |
Closest papers by meaning
- Trigger Hunting with a Topological Prior for Trojan Detection (ICLR 2022 · similarity 0.90)
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 2024 · similarity 0.87)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.86)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.86)
- DeDe: Detecting Backdoor Samples for SSL Encoders via Decoders (CVPR 2025 · similarity 0.85)
- Protecting Model Adaptation from Trojans in the Unlabeled Data (AAAI 2025 · similarity 0.85)
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.
