"Demystifying Privacy Preserving Computing" by Tejas Chopra (Strange Loop 2022)

### Overview In this talk from Strange Loop 2022, Tejas Chopra, a senior software engineer at Netflix, demystifies privacy-preserving computation. The presentation explores how modern systems can perform meaningful computation and machine learning over private data without sacrificing confidentiality. Chopra outlines the tension between full privacy (no data sharing, zero utility) and full utility (unrestricted data exposure, zero privacy), motivating the need for middle-ground techniques that allow organizations to compute over sensitive data safely in domains like healthcare, cloud computing, and blockchain.

### Key Topics Chopra categorizes privacy-preserving computation into four foundational paradigms: - **Federated Learning:** Pushing machine learning models to edge devices (e.g., Apple Siri, Google keyboard autocompletion) rather than aggregating raw user data centrally. Devices train models locally and return only updated parameters. - **Differential Privacy:** A formal mathematical framework that quantifies privacy loss (measured by epsilon) by ensuring the presence or absence of a single individual's record does not significantly alter aggregate query results, typically achieved by injecting calibrated noise. - **Secure Multi-Party Computation (SMPC):** Protocols enabling multiple parties to jointly evaluate functions across distributed secret shares without revealing private inputs to any participant, often leveraging Shamir's Secret Sharing. - **Homomorphic Encryption:** Cryptographic schemes allowing mathematical operations directly on encrypted ciphertexts, producing an encrypted result that decrypts to the plaintext answer.

### Technical Details - **Federated Learning Trade-offs:** While compliant with privacy regulations like GDPR, it faces device-level compute disparities and "leaky abstractions," where adversaries can reverse-engineer models or infer sensitive data by sniffing updated parameters. - **Differential Privacy & Epsilon:** Differential privacy bounds individual impact through an epsilon (ε) threshold. While used by organizations like the U.S. Census Bureau (2020), Google Chrome, and Microsoft, researchers identified issues with transparency, citing cases where Apple utilized very high epsilon values that leaked more privacy than anticipated. - **SMPC & Shamir's Secret Sharing:** SMPC uses polynomial interpolation where a degree-t polynomial hides a secret at $x=0$. It requires at least $t+1$ points to reconstruct the secret, making it information-theoretically secure against unbounded or quantum computing attacks. However, SMPC suffers from high communication latency among nodes and vulnerability to colluding or malicious parties. - **Homomorphic Encryption Challenges:** Schemes span partial (limited operations), somewhat (limited depth due to accumulating noise), and fully homomorphic encryption (FHE). FHE focuses primarily on addition and multiplication, analogous to universal logic gates. IBM's HElib historically ran a trillion times slower than plaintext (reduced to ~75x slower), while Microsoft developed the SEAL library. Noise accumulation remains the central technical challenge during multiplication, requiring noise-resetting mechanisms. - **Blockchain and Zero-Knowledge Proofs:** During the Q&A, Chopra highlights privacy gaps in public ledgers, noting privacy blockchains like Zcash transitioning from ZK-SNARKs (which required complex trusted setup phases) toward ZK-STARKs to preserve transactional confidentiality.

### Takeaways Privacy-preserving technologies are not a substitute for core security and authentication; they operate as authorization and data-protection layers. The current adoption of SMPC and FHE is constrained by high computational overhead, communication latency, poor interoperability, and lack of standardized benchmarking across libraries. The future of privacy-preserving systems lies in hybrid architectures—combining SMPC for distributed processing, FHE for encrypted computation, and differential privacy to quantify disclosure risks.