Deep Leakage — Reconstructing Data From Model Gradients

Why it matters

Federated learning promised privacy. Deep leakage shows it's not automatic. Understanding shapes protections.

The architecture

Gradients encode information about training data.

Iterative reconstruction: guess input, compute gradient, adjust to match observed gradient.

How it works end to end

Attacks recovered high-resolution images and text from single gradient updates.

Defenses: differential privacy noise added to gradients. Secure multi-party aggregation. Gradient clipping + noise.