Selective Data Protection for Energy-Efficient Deep Neural Network Inference in Intermittent IoT Systems

Intermittent computing systems allow devices to operate under unstable and weak power conditions. When a device performing long tasks is interrupted due to power instability, a checkpoint mechanism is used to save data to non-volatile memory before the system shuts down, allowing computation to resume when power is restored. However, power failures can leave data stored in nonvolatile memory vulnerable to tampering during the downtime. To address this, recent research has proposed encrypting and verifying data to protect it and prevent replay attacks that could disrupt program execution and deep inference sequences.