SPIRT: A Fault-Tolerant and Reliable Peer-to-Peer Serverless ML Training Architecture

Distributed machine learning (DML) has evolved rapidly, driven by the need to train large-scale models on massive datasets. Parameter server (PS) architectures, exemplified by Google’s TensorFlow, facilitate centralized coordination but suffer from single points of failure and scalability issues. Conversely, P2P architectures distribute data and computation across nodes, enhancing robustness and privacy. Recent advances include federated learning and decentralized algorithms, which address data privacy and system resilience. However, integrating these with serverless computing remains challenging due to statelessness and communication overhead. Existing works like LambdaML and SMLT leverage databases for communication but face latency issues. The gap lies in achieving efficient, fault-tolerant, and secure P2P training within a serverless framework, which this work aims to fill.

Current P2P ML systems struggle with high communication latency, node failures, and security threats such as Byzantine attacks. The stateless nature of serverless platforms complicates synchronization and model consistency. Traditional approaches rely on external communication channels, increasing delays and vulnerability. Ensuring continuous training despite node failures, malicious nodes, and communication bottlenecks remains unresolved. Moreover, existing robust aggregation methods are computationally intensive and not optimized for database-internal execution, limiting scalability. Addressing these issues requires a novel architecture that combines in-database computation, automated orchestration, and security measures to enable reliable, efficient, and scalable P2P ML training.

The core innovation is embedding model and gradient updates directly within RedisAI, eliminating external data transfers and reducing latency. The architecture employs AWS Step Functions for seamless workflow automation, enabling dynamic management of training epochs, node addition, and fault recovery. Integration of robust aggregation algorithms like KRUM enhances security against Byzantine nodes. The system supports secure peer authentication via RSA encryption and digital signatures, ensuring data integrity. Data redistribution mechanisms handle node failures gracefully, maintaining training progress. This holistic design combines database-centric computation, serverless orchestration, and security, setting a new standard for decentralized ML frameworks.

• �� Initialization: Nodes generate RSA key pairs, authenticate via AWS KMS, and exchange encrypted credentials. • Data Partitioning: Datasets stored in S3 are segmented into batches assigned per node. • Gradient Computation: AWS Lambda functions compute gradients in parallel, storing results in Redis. • Gradient Averaging: Lua scripts within Redis perform in-database averaging, reducing external communication. • Synchronization: Nodes notify completion via SQS queues; a Lambda function triggers aggregation once all are ready. • Robust Aggregation: Pull gradients from active nodes, apply KRUM to exclude outliers, and produce a final gradient. • Model Update: RedisAI updates model parameters directly in Redis. • Convergence Check: Periodic Lambda functions evaluate model convergence criteria. • Fault Tolerance: Detect inactive nodes, redistribute data, and adjust training flow dynamically.

Experiments utilized MNIST dataset, training ResNet-18, MobileNetV3, and DenseNet-121 models across 10-50 nodes. Metrics included training time, accuracy, robustness under Byzantine attacks, and fault recovery speed. Baseline comparisons involved traditional parameter server and semi-decentralized schemes. Hyperparameters such as batch size, learning rate, and aggregation thresholds were tuned for optimal performance. Ablation studies assessed the impact of database-internal updates and robust aggregation. Results consistently showed 82% reduction in update time, >98% accuracy under attack, and rapid recovery from simulated node failures, validating the architecture’s efficiency and resilience.

The system achieved an 82% decrease in model update time compared to conventional distributed ML setups. Accuracy remained above 98% even under Byzantine attack simulations, demonstrating robust defense. Fault injection tests showed the system could detect and recover from node failures within seconds, maintaining training continuity. Database-internal updates significantly reduced communication latency, leading to faster convergence. Scalability tests indicated stable performance up to 50 nodes, with potential for larger deployments after further optimization. These results confirm the architecture’s effectiveness in real-world distributed AI scenarios.

This architecture is ideal for privacy-sensitive applications like healthcare and finance, where data cannot be centralized. It supports edge computing environments, enabling collaborative training across geographically dispersed devices. The system’s fault tolerance and security features make it suitable for mission-critical AI deployments requiring high reliability and data integrity. Additionally, it can facilitate federated learning scenarios, reducing communication costs and enhancing privacy protections.

The reliance on cloud services like AWS may limit deployment flexibility in private or on-premise settings. Handling extremely large models remains challenging due to Redis storage constraints. The robustness against sophisticated Byzantine attacks needs further validation in diverse real-world environments. Computational costs associated with cryptographic operations and orchestration layers could impact efficiency at scale. Future work should focus on optimizing resource usage and extending support for larger models and more complex attack scenarios.