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.