Lecture, Keynote
Raluca Ada Popa
- Problem Context: Traditional Deep Packet Inspection (DPI) middle boxes inspect packet payloads to detect intrusions, prevent data exfiltration, and enable parental filtering, but these functions are thwarted by widespread HTTPS encryption.
- Current Insecure Solution: To bypass encryption, organizations currently perform Man-in-the-Middle (MITM) attacks using fake certificates, which decrypts traffic at the middle box and destroys end-to-end privacy guarantees.
- Blindbox Objective: The system aims to enable DPI middle boxes to inspect encrypted payloads without ever decrypting the traffic, maintaining strong privacy while providing security functionality.
- Blindbox Architecture: The system introduces "Blindbox HTTPS," where user traffic is encrypted via standard SSL and then tokenized; keywords are encrypted into tokens and sent to the middle box for inspection without decryption.
- Encryption Mechanism: Blindbox utilizes a novel searchable encryption scheme that combines the high security of randomized encryption (resisting frequency analysis) with the speed of deterministic encryption (enabling indexing), using AES hardware instructions.
- Key Management: Encrypted rules are generated via a Secure Multi-Party Computation (MPC) protocol during the SSL handshake, allowing the middle box and user to agree on encrypted values without either party revealing their private keys or rules.
- Threat Model: The system assumes a powerful threat model where the middle box administrator has access to all traffic but must learn only whether a specific malicious string matches, gaining no other information.
- Performance Metrics: Middle box detection performance is competitive with Snort, though end-to-end page load times increase by 15% to 100% depending on connection duration due to SSL handshake overhead.
- Long-Lived Connection Viability: While short-lived connections suffer from handshake latency, the performance impact is negligible for long-lived enterprise-to-cloud connections.
- Functional Scope: The system currently supports equality checks for watermarks, exfiltration detection, and parental filtering, with a secondary protocol supporting complex regular expressions.
- Broader Vision (Computing on Encrypted Data): Blindbox is part of a trend toward systems where data remains encrypted throughout storage, transit, and computation, moving beyond traditional "encrypt at rest/transit" models.
- Related Systems & Impact:
- CryptDB: An encrypted database system capable of processing SQL queries, showing a 27% overhead on TPCC benchmarks compared to MySQL; used by SAP, MIT Lincoln Lab, and Google's BigQuery.
- Milter: A web server platform ensuring confidential data is processed in encrypted form; currently deployed in a Boston hospital for medical applications.
- Location & ML Applications: The technology supports private location aggregation for traffic mapping and machine learning classification over encrypted data.
- Industry Adoption: Companies such as Sky High and Microsoft (via "Always Encrypted") are leveraging these encryption-on-computation techniques, signaling a shift toward practical confidential computing.