TL;DR
This paper introduces Noise Information Secretly Sharing (NISS), a novel algorithm for Federated Learning (FL) that recovers the model accuracy typically lost when using Differential Privacy (DP). By having clients secretly share negated versions of their DP noise with each other, the noise "cancels out" at the server level while remaining high enough at the client level to prevent data leakage, resulting in an average accuracy boost of 21%.
What Problem Does This Paper Address?
Federated Learning allows multiple devices (clients) to train a shared model without ever sending their raw data to a central server. However, researchers have found that private data can still be "leaked" or reconstructed from the gradients clients send to the server.
To stop this, practitioners use Differential Privacy (DP), which involves adding random Gaussian noise to the gradients before uploading them. While DP protects privacy, it creates a "utility-privacy trade-off": the noise that hides the data also confuses the model, often dropping accuracy by 10–30%. This makes DP impractical for real-world systems that require high precision. The authors aim to achieve "Gain without Pain"—maintaining the privacy of DP without the accuracy penalty.
Methods
The authors propose the NISS algorithm, which combines elements of DP and Secret Sharing (SS). The process works as follows:
- Noise Generation: Each client generates Gaussian noise ($n^k$) to protect their local gradient.
- Secret Sharing: Instead of just adding the noise and uploading, the client splits a negated version of that noise into multiple shares and sends these shares to a set of "neighbor" clients.
- Distortion for Security: To prevent a malicious server from colluding with other clients to "undo" the noise, clients multiply the received shares by a random distortion variable ($s_i$). This ensures that even if some neighbors are untrustworthy, the client's privacy remains intact.
- Uploading: Each client uploads their gradient plus their own noise plus the negated noise shares they received from others.
- Aggregation: When the Parameter Server (PS) sums up all the updates, the original noises and the negated shares mathematically offset each other. If clients are trustworthy, the noise cancels out perfectly ($V=0$), leaving a clean, accurate aggregate gradient.
Datasets
The algorithm was evaluated using two standard computer vision benchmarks:
- MNIST: A dataset of 70,000 grayscale images of handwritten digits.
- CIFAR-10: A dataset of 60,000 color images across 10 different object classes (e.g., airplanes, cats, dogs). The researchers tested these across both IID (Independent and Identically Distributed) and non-IID settings to simulate real-world data heterogeneity.
Key Insights and Ideas
- The "Offsetting" Magic: The core "aha" moment is that noise doesn't have to be permanent. By making noises negatively correlated across the network, they can protect individual privacy (where the noise is high) but vanish during aggregation (where the sum is what matters for the model).
- Better Privacy than Standard DP: Surprisingly, NISS can provide better privacy than standard DP. In standard DP, the noise level is limited by the accuracy drop we are willing to tolerate. In NISS, because the noise cancels out, clients can inject much larger amounts of noise locally, making it even harder for an attacker to perform gradient leakage attacks.
- Stealthy Defense: The use of the distortion factor $\tau$ allows the system to be "tunable." If the environment is highly untrustworthy, clients can increase $\tau$, which keeps more noise in the final aggregate to ensure safety, albeit with a slight hit to accuracy.
Key Results
- Accuracy Recovery: NISS improved model accuracy by an average of 21% compared to standard DP-FedAVG.
- Near-Baseline Performance: When clients are trustworthy, NISS achieves accuracy levels nearly identical to FedAVG (the standard algorithm with no privacy noise at all).
- Robustness to Attacks: Using the Deep Leakage from Gradients (DLG) metric, the authors showed that while standard DP and FedAVG allowed for partial or total image reconstruction by an attacker, NISS effectively blocked these attacks, resulting in much higher "DLG loss" (meaning the attacker failed to recover meaningful info).
- Non-IID Success: On the MNIST dataset with non-IID data, NISS increased test accuracy by 30–40% over standard DP methods.
Limitations and Future Work
- Communication Overhead: NISS requires clients to communicate with each other (peer-to-peer) to share noise shares. In a cross-device setting (like mobile phones), this peer-to-peer communication can be expensive or unreliable.
- The Tracker Server: The system relies on a "tracker server" to help clients find neighbors. While the authors state this server doesn't see private data, it represents an additional piece of infrastructure.
- Collusion Limits: While NISS handles some level of collusion, the authors note a theoretical trade-off: if a very high percentage of clients collude with the server, the accuracy gains of NISS diminish as the system must revert to standard DP levels to maintain privacy.
- Future Work: The authors suggest exploring NISS in more complex, large-scale IoT environments where device trust can be modeled more dynamically.
Summary updated