Idea
A privacy-preserving split learning platform that secures training labels and reduces costs for enterprises handling sensitive ML data
Research Paper
Core Innovation
This paper introduces SplitHappens, which uniquely combines Function Secret Sharing with U-shaped Split Learning to keep training labels secret from servers. It addresses advanced threats like model inversion and label inference attacks while reducing training time and communication overhead. This approach improves security without sacrificing model accuracy.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing demand for secure and efficient machine learning in regulated industries and cloud environments.
Potential Customers & Pain Points
- Enterprises Handling Sensitive Data Needing Enhanced Privacy
- AI Developers Concerned About Label Leakage and Model Inversion Attacks
- Cloud Service Providers Seeking Efficient Secure ML Training
Business Model
Subscription-based SaaS platform offering secure split learning APIs and enterprise integration support
Competitive Landscape
- SplitML
- CrypTFlow
- Duality Technologies
Implementation Challenges
- Complexity of integrating FSS with existing ML pipelines
- Adoption resistance due to new cryptographic methods
- Performance trade-offs in large-scale deployments
Validation Strategy
- Develop prototype integrating FSS with split learning
- Pilot with select enterprises handling sensitive data
- Measure privacy improvements and training efficiency gains
Research Paper Overview
Split Happens: Combating Advanced Threats with Split Learning and Function Secret Sharing
Summary
This paper presents SplitHappens, a novel approach combining Function Secret Sharing (FSS) with U-shaped Split Learning (SL) to enhance data privacy in machine learning. It improves security guarantees by keeping training labels secret from servers, addresses modern attack vectors like model inversion and label inference, and reduces training time and communication costs while maintaining accuracy.