Authors: Mohit Sewak (Microsoft), Sanjay K. Sahay (BITS Pilani), Hemant Rathore (BITS Pilani)



Motivated by the advancements made by the recently proposed DRo algorithm to uplift the performance of data scarce Deep Learning malware detector for edge, we propose an adaptive and efficient system for hybrid edge-cloud detection and forensics, named GreenForensics. The proposed adaptive enhancement, makes the system more suitable for devices with custom battery-performance optimization mandates like tablets and laptops. Further, the enhancements offer various discrete and continuous controls for influencing the detection coverage and model footprints in real time. To further enhance the detection efficiency and making the detection resilient to adversarial-attacks, the proposed system can work with adversarial-DL immune algorithms. In the experiments conducted, GreenForensics was able to significantly outperform even the best baseline deep architectures and improved the detection and forensics robustness by up to 100% and performance by up to 40%. This gains further significance as the incumbent baseline DL architecture had up to 6700% higher neural inference complexity, and had its performance and robustness benchmarks had remained unchallenged for a long time.