Abstract
Federated Learning (FL) combined with Homomorphic Encryption (HE) offers a robust privacy-preserving training paradigm by allowing mathematical operations on encrypted data. However, the integration of HE often leads to a "triple threat" of challenges: significant computational overhead, a lack of transparency for auditing malicious updates, and difficulties in assessing participant contributions for fair rewards. This paper introduces the FLSSM (Federated Learning Storage Security Model) framework, which utilizes hierarchical parallel aggregation, secret sharing for conditional auditing, and trusted timestamps for incentive distribution to overcome these bottlenecks.
The Triple Threat of Encrypted Federated Learning
While Federated Learning allows multiple parties to train a model without sharing raw data, gradients are still vulnerable to inference attacks. Homomorphic Encryption (HE) is the standard defense, but it introduces three critical problems that hinder practical deployment:
- Computation Efficiency: HE operations are computationally intensive, creating a massive bottleneck at the central aggregation server as the number of participants grows.
- Attack Tracing: Encrypted models act as a "black box." A central server cannot inspect updates to determine if a participant is submitting a poisoned model or a backdoor attack.
- Contribution Assessment: Without the ability to inspect gradients, it is nearly impossible to determine which participants provided high-quality data, making fair compensation difficult.
The FLSSM Architecture
To address these issues, the authors propose a multi-layered framework consisting of three primary modules designed to balance security with operational efficiency.
1. Hierarchical Aggregation Mechanism (HAM)
To solve the computation bottleneck, FLSSM employs a "sharding" strategy. The global model is split into slices, and different edge aggregation nodes process these slices in parallel using the CKKS encryption algorithm. This distributed approach prevents a single-server bottleneck and ensures that global model updates remain stable even as the network scales.
2. Model Access Control Mechanism (MACM)
To enable auditing without compromising privacy, FLSSM uses Shamir’s Secret Sharing. The decryption key is divided among trusted "supervise nodes." If the global model's performance drops—signaling a potential attack—these nodes can collaborate to reconstruct the key and audit the suspicious local model. These shares are further secured using Attribute-Based Encryption (ABE) to ensure only authorized supervisors can participate in the reconstruction.
3. Incentive Mechanism based on Trusted Time Intervals (IMTTI)
Since the server cannot inspect data quality directly, it uses Trusted Timestamps from an external authority (TSA) as a proxy for effort. By verifying the time a node spent on local training, the system can calculate rewards fairly. This assumes that nodes dedicating more verifiable resources are likely contributing more significantly to the model's progress.
Experimental Evaluation
The model was evaluated on standard computer vision benchmarks, CIFAR-10 and Fashion-MNIST, using a Dirichlet distribution to simulate real-world Non-IID (non-independent and identically distributed) data environments.
- Efficiency: The HAM approach using 10 edge nodes significantly outperformed standard central CKKS aggregation. Aggregation times remained stable even as the number of training nodes increased.
- Security: In scenarios where 10-20% of nodes performed sign-flipping attacks, the MACM successfully identified and blacklisted attackers, allowing the global model to recover its accuracy.
- Incentives: The IMTTI successfully identified nodes attempting to report false training times, ensuring that rewards were distributed based on verified effort.
Conclusion and Future Work
FLSSM demonstrates that privacy in Federated Learning does not have to come at the cost of efficiency or accountability. By shifting from a centralized to a hierarchical architecture and implementing "conditional transparency" through secret sharing, the framework provides a viable path for secure, large-scale FL. Future research will focus on refining the "time-as-quality" assumption and reducing the communication overhead associated with sharded encrypted updates.
Summary updated