PMC: A Privacy-preserving Deep Learning Model Customization Framework for Edge Computing
summary
PMC extracts channel-wise Gaussian parameters from intermediate activations on-device, adds noise for differential privacy, uploads only those stats, and reweights cloud data to retrain a customized CNN without raw data. Experiments on public datasets show large accuracy gains, especially at high pruning ratios, beating adaptation baselines and matching privacy-free EFI, and a user study shows the stats distinguish users from different environments; best practice is selecting 50–100 channels by mean activation.
AI-written summary · about
related talks
31mScaling privacy embedded distributed ML systems
14mEstablishing Privacy Metrics for Genomic Data Analysis
29mCloud and Edge: Price, Performance and Privacy Considerations in IOT
44mPersonalized Federated Training of Diffusion Models with Privacy Guarantees
31mIs your model private?
33mPrivacy of Decentralized Machine Learning
17mPrivate Tuning of LLMs in Practice: From VaultGemma to Custom Fine-Tuning
16mPrvTel: Lightweight Models for Private and Accurate Telemetry Data Retention