Closing Remarks and Q&A
Computer Vision over Homomorphically Encrypted Data
CVPR 2025 Tutorial
June 12, 2025
Tutorial Summary
- Data security is critical for real-world computer vision systems.
- End-to-end encrypted computer vision systems with FHE are feasible.
- Data security can be realized.
- Offers strong post-quantum security guarantees.
- Naive application of FHE for computer vision suffers from prohibitive computational costs.
- Rethinking neural architecture designs is critical for efficiency.
- Can get orders of magnitude speed up.
Exciting research at the intersection of AI and Cryptography.
From Pixels to Decisions—Encrypted End-to-End
Feasible, but work remains to make it practical.
Emerging Problems
FHE Enables computations on untrusted environments
Checking integrity of biometric operations under Encryption
We trust, BUT we do not verify[BHV+21] Fast and accurate likelihood ratio-based biometric verification secure against malicious adversaries
Going Beyond Data Privacy
What are we trying to protect in AI?
- $x$: images, audio, video, text
- $f$: parameters, functional form
Function Privacy
- Protect intellectual property.
- Prevent attacks against model.
- Prevent leakage of training data.
- Comply with industry security standards.
Story So Far...
Co-designing AI and FHE architectures is critical for efficiency.
Missing Piece in the Puzzle
Hardware Accelerators for FHE
Homomorphic Processing Unit
Concluding Remarks
Key Takeaways
- Secure AI is achievable with Fully Homomorphic Encryption (FHE).
- Efficient and specialized architectures are crucial for practical encrypted inference.
- Real-world applications, like secure facial recognition, demonstrate feasibility today.
Next Steps
- Explore our open-source code & tutorials.
- Stay connected through community resources (fhe.org Discord).
- Collaborate on future research challenges!
We appreciate your interest. Let's advance secure computer vision together.