VeriSplit: Secure and Practical Offloading of Machine Learning Inference across IoT Devices

VeriSplit is a framework that offloads machine learning inference from IoT devices to locally available devices; prototype evaluation shows it can reduce inference latency by 28%–83% compared with local computation. The framework uses masking techniques to protect data privacy and model confidentiality, and uses a commitment-based verification protocol to guarantee inference integrity. Unlike existing schemes that rely on finite field computation, it does not interfere with the floating-point operations of hardware accelerators and requires no modification to existing models.