Federated Neuro-Symbolic Learning for Privacy-Preserving Intelligent Edge Systems

Abstract

 Intelligent Edge Systems – smart cameras, wearable technology, connected cars and industrial sensors – are creating a growing need to learn from local data at the edge. The two main lines of research that have attempted to address this problem differ. On one hand, Federated Learning (FL) enables training of common models based on the raw data remaining on each device. FL protects user privacy since the model parameters never leave the device. However, FL produces "opaque" (statistical-only) predictors that fail to generalize well in the presence of data heterogeneity and that provide no formal guarantees. On the other hand, Neuro-Symbolic (NeSy) learning combines neural perception (e.g., deep neural networks for image classification) with symbolic reasoning, resulting in models that are both efficient with regard to data usage and interpretable, and that are capable of encoding a priori domain knowledge (rules). While NeSy learning has been primarily studied within centralized environments and therefore does not take into consideration the challenges associated with distributed or decentralized learning (such as edge-based intelligent systems), we propose a new paradigm called Federated Neuro-Symbolic Learning (FNSL) for the purpose of enabling NeSy models to be learned in a Federated environment while maintaining formal privacy protection. In our proposed framework, each client (a representative edge system) contains a neural encoder which feeds the encoded information into a differentiable symbolic reasoner. Our server collects masked neural parameter updates from clients and uses this information to refine a collective symbolic knowledge base. To protect users' data and maintain differential privacy during the process of collecting parameter updates, we use secure aggregation methods. Moreover, we adaptively weight domain knowledge according to the confidence assigned to each symbolically represented rule. We demonstrate through experiments using image, sensor and tabular benchmark datasets and non-IID partitioning schemes that FNSL achieves improvements of 4-7 percentage points in accuracy compared to state-of-the-art FL baselines. Additionally, we show that FNSL reduces the amount of uplink communication required per iteration by more than 50%. Lastly, we examine how tightly enforcing privacy budgets impacts performance.