local differential privacy in fl

**Local Differential Privacy (LDP) in FL** is a **stronger privacy model where each client adds noise to their gradient update BEFORE sending it to the server** — the server never sees the true gradient, providing privacy even against an untrusted server. **LDP vs. Central DP** - **Central DP**: Server receives true client updates, then adds noise — requires trusting the server. - **LDP**: Each client adds noise locally before sending — privacy holds against any server, malicious or honest. - **Noise Level**: LDP requires $sqrt{n} imes$ more noise than central DP for the same privacy guarantee ($n$ = number of clients). - **Utility**: LDP has significantly lower model accuracy than central DP — much more noise needed. **Why It Matters** - **Zero Trust**: Privacy is guaranteed even if the server is compromised or malicious. - **Regulatory**: Some regulations require data protection against the service provider — LDP satisfies this. - **Practical Trade-Off**: LDP privacy comes at a steep accuracy cost — only viable with many clients. **LDP in FL** is **privacy without trusting anyone** — each client protects their own data locally, eliminating the need to trust the aggregation server.