IIT Madras' FedBNN Shrinks On-Device AI Models 32x Without Losing Accuracy

Researchers at IIT Madras and UIUC have proposed FedBNN , a framework that trains binary neural networks within the federated learning loop. The paper reports reductions of up to 58 times in floating-point operations and 32 times in model memory, while retaining accuracy close to full-precision federated baselines across several image-classification benchmarks.

Federated learning reduces the movement of raw data by training across participating devices, but the resulting model must still fit each device’s memory, compute, and energy budget. FedBNN targets that deployment constraint by representing weights and activations with one bit during inference. Model updates can still reveal information, so secure aggregation, differential privacy, and related protections remain separate concerns.

Full-precision training leaves an inference problem

Common FedAvg deployments train models whose weights are stored as 32-bit floating-point values. That format works well on servers, but it can strain phones, sensors, and microcontrollers that must run the model repeatedly under tight power and memory limits.

Binary neural networks compress weights and activations to values such as plus one and minus one. Packed binary kernels can replace floating-point multiplication with XNOR operations, which compare bits, and population-count operations, which count matching results. Post-training binarization often loses substantial accuracy because rounding a trained floating-point model onto that two-value grid introduces large errors.