Shafi Goldwasser Provides 'A Cryptographic Perspective on Trustworthy AI'

2026-03-03 · Source: MIT CSAIL · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Shafi Goldwasser, a pioneer in modern cryptography, presented a cryptographic perspective on trustworthy AI, emphasizing the rapid adoption of machine learning models like large language models (LLMs) and the critical need for their reliability and trustworthiness. She outlined a cryptographic recipe for addressing AI trust issues: modeling adversaries, defining sufficient security, proposing solutions, and providing proofs, often based on computational hardness assumptions. Goldwasser detailed five key challenges: privacy in training data, verification of model properties (correctness, fairness, robustness), robustness against adversarial data shifts and backdoors, alignment with human values, and ownership to prevent model stealing. She highlighted the use of homomorphic encryption and trusted hardware for privacy, interactive proofs for verification, and cryptographic techniques like time-lock puzzles for alignment and robustness against backdoors, noting that simple filtering is often insufficient for safety.

Key takeaway

For AI scientists and developers building or deploying machine learning models in sensitive applications, you should adopt a security-first mindset by integrating cryptographic principles. Focus on defining adversarial models and proving the trustworthiness of your AI systems, rather than relying solely on empirical accuracy. Consider using techniques like homomorphic encryption for data privacy during training and developing self-proving models to ensure verifiable correctness, especially for critical outputs. Be aware that simple filtering mechanisms are often inadequate for ensuring model alignment and safety against sophisticated "jailbreak" attacks.

Key insights

Cryptography offers a robust framework for defining, achieving, and proving trustworthiness in AI systems against adversarial threats.

Principles

• Trust in AI requires adversarial modeling and provable security.
• Worst-case guarantees are superior to average-case accuracy for critical AI applications.
• Simple input/output filtering is insufficient for AI safety and alignment.

Method

The cryptographic recipe for trustworthy AI involves modeling adversaries, defining security, proposing solutions (e.g., homomorphic encryption, interactive proofs), and providing mathematical proofs of security under computational assumptions.

In practice

• Utilize homomorphic encryption for privacy-preserving ML training.
• Implement verifier-guided reinforcement learning for self-proving models.
• Post-process models to mitigate undetectable backdoors.

Topics

• Trustworthy AI
• Cryptographic AI
• Homomorphic Encryption
• AI Model Robustness
• LLM Alignment

Best for: AI Scientist, AI Researcher, Research Scientist, Machine Learning Engineer

Related on AIssential

• Model Stealing Through the Lens of Model Multiplicity — High-fidelity stolen models may not be functionally equivalent to originals due to model multiplicity.
• Dapr 1.18 Introduces Verifiable Execution, Bringing Cryptographic Trust to AI Agents and Workflows — Dapr 1.18 introduces Verifiable Execution to establish cryptographic trust and provenance for AI agents and distributed workflows.
• Are Safety Guarantees in Neural Networks Safe? How to Compute Trustworthy Robustness Certifications — A new apothem measure enables efficient, trustworthy robustness certifications for neural networks, outperforming volume-based methods.
• Private analytics via zero-trust aggregation — A zero-trust private analytics solution combines one-shot cryptographic aggregation with TEEs for robust, multi-layered data protection.
• TrustedARI: Towards Trust-Native Agentic Routing Infrastructure for Agentic AI — TrustedARI secures AI agent interactions with external services via novel cryptographic protocols, ensuring privacy and verifiable transactions.
• Is AI Trustworthy? — AI's biggest opportunity lies in building trust for high-stakes enterprise automation, not just in model intelligence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MIT CSAIL.