Patch-Level DP

🔒 Transactions on Machine Learning Research (TMLR) 2026

Overview

What if you could get stronger privacy guarantees for free — just by doing something you already do during training?

This work reveals that random cropping, a ubiquitous data augmentation in computer vision, naturally amplifies patch-level differential privacy. No extra computation, no architectural changes, no overhead — just stronger privacy from a technique already in your pipeline.

Key Contributions

• 🆓 Privacy for free — leverages standard random cropping to amplify patch-level DP guarantees
• 🧩 Patch-level analysis — provides fine-grained privacy at the image patch level
• ⚡ Zero overhead — no additional computation or model changes required
• 🔧 Drop-in compatible — works with any vision pipeline that already uses random cropping

Why It Matters

Privacy-preserving ML often comes at a steep cost in complexity or performance. This work shows that in computer vision, you may already be getting more privacy than you think — and formalizes exactly how much.

(Durmaz et al., 2026)