Secure human action recognition by encrypted neural network inference.

Area of Science:

• Computer Vision
• Cryptography
• Health Informatics

Background:

• Advanced computer vision offers real-time home monitoring for aging in place, detecting critical events like falls, seizures, and stroke symptoms.
• Affordable webcams and cloud computing for machine learning present significant social benefits but face practical deployment barriers due to privacy concerns.

Purpose of the Study:

• To propose a novel strategy using homomorphic encryption to address privacy concerns in computer vision-based home monitoring.
• To develop a secure inference protocol capable of distinguishing falls from daily living activities while ensuring data confidentiality.

Main Methods:

• Implementation of a homomorphic encryption strategy for secure action detection in home monitoring.
• Development of a secure inference protocol to analyze video data without compromising privacy.

Main Results:

• The secure inference protocol achieved 86.21% sensitivity and 99.14% specificity in distinguishing falls from activities of daily living.
• Average inference latency was 1.2 seconds (small neural nets) and 2.4 seconds (large neural nets).
• The method demonstrated a 613x speedup over latency-optimized LoLa and a 3.1x throughput increase compared to nGraph-HE2.

Conclusions:

• Homomorphic encryption effectively resolves the privacy dilemma in computer vision-based home monitoring.
• The proposed secure inference method enables practical, privacy-preserving solutions for supporting aging in place.