Research map / Adversarial robustness and security / Attacks on text and graphs
Anti-Neuron Watermarking: Protecting Personal Data against Unauthorized Neural Networks
Zihang Zou, Boqing Gong, Liqiang Wang
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 |
|---|---|---|---|
| add LCT watermark to user images (cosine similarity of inferred signature) | ResNet50 on Tiny ImageNet · Tiny ImageNet (0.1% watermarked) | cosine similarity: 0.045 -> 0.999 | Improved |
| add LCT watermarking vs dataset tracing | Tiny ImageNet · Tiny ImageNet (0.1% watermarked) | cosine similarity after training: -0.005 -> 0.999 | Improved |
| add LCT watermarking to user data | ResNet50 · Tiny ImageNet user data | memorization value estimate: 26.8% -> 34.8% | Improved |
| reduce watermarked samples from 5 to 1 (ablation) | LCT watermarking · Tiny ImageNet gradient search | signature recovery: 60.99 matched -> 17.0 missed | Hurt |
| vary number of watermarked samples 10,000 to 5 (ablation) | LCT watermarking · Tiny ImageNet | model accuracy: 55.6 to 54.6 | No clear effect |
| vary the watermark signature (60 to 300) (ablation) | LCT watermarking · Tiny ImageNet | signature recovery: inferred 59.5, 120.2, 178.8, 242.1, 306.1 | No clear effect |
| vary classifier architecture (Alex, VGG, ResNet, Wide Res, Dense) (ablation) | LCT watermarking · Tiny ImageNet | signature recovery: inferred 56.9, 52.8, 59.5, 58.4, 58.8 | No clear effect |
| increase model capacity (ResNet18 to ResNet152) (ablation) | LCT watermarking · Tiny ImageNet | loss minimum sharpness | Improved |
| train with color jitter augmentation including hue | LCT watermarking · Tiny ImageNet | loss gap clean vs watermark: inferred 64.8 for signature 60 | Hurt |
| strong augmentation on CUB-200-Birds appendix | LCT watermarking · CUB-200-Birds | signature recovery | Hurt |
| more users with the same LCT (ablation) | LCT watermarking · Tiny ImageNet 1,000 users | matching accuracy: almost all correct at 20% users, drops significantly with same LCT | Hurt |
| user-specific random LCTs instead of the same LCT (ablation) | LCT watermarking · Tiny ImageNet 1,000 users | matching accuracy: above 80% with all users watermarked | Improved |
Closest papers by meaning
- Watermarking Deep Neural Networks with Greedy Residuals (ICML 2021 · similarity 0.88)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.88)
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.87)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.86)
- Resource Efficient Deep Learning Hardware Watermarks with Signature Alignment (AAAI 2024 · similarity 0.86)
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · 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.
