Homomorphic Encryption vs Secure Multi-Party Computation for Inter-Agency AI Queries

Homomorphic Encryption (HE) excels at enabling computation on encrypted data by a single, untrusted party because it allows mathematical operations to be performed directly on ciphertexts. For example, a central agency could run an AI inference model on encrypted citizen data from multiple departments without ever decrypting the raw inputs, resulting in a computational overhead that can be 1,000x to 100,000x slower than plaintext operations, depending on the scheme (e.g., CKKS for approximate arithmetic vs. TFHE for fast bootstrapping). This makes HE ideal for low-complexity, high-privacy tasks like encrypted keyword search or simple risk scoring.

Secure Multi-Party Computation (MPC) takes a different approach by distributing secret-shared data among multiple non-colluding servers that jointly compute a function. This results in a fundamentally different trade-off: high communication overhead between parties but significantly lower computational latency per node compared to HE. For instance, an inter-agency query to identify benefit fraud could be completed in seconds with a 3-party replicated secret sharing protocol, but it requires constant high-bandwidth communication and the strict assumption that a majority of the servers remain honest.

The key trade-off: If your priority is a zero-trust architecture where data remains encrypted even during computation by a single processor, choose Homomorphic Encryption. If you prioritize lower latency and complex computation across a small, semi-trusted federation of agencies, choose Secure Multi-Party Computation. For real-time law enforcement queries, the communication rounds in MPC often prove more practical than the extreme computational latency of HE, while HE's non-interactive nature is superior for asynchronous, long-running social service analytics.

Homomorphic Encryption (HE) excels at protecting data-in-use during computation because it allows a single, untrusted party to perform calculations directly on ciphertexts without ever decrypting them. For example, a central analytics node could compute aggregate statistics over encrypted tax records from multiple agencies. However, this mathematical purity comes at a steep price: fully homomorphic encryption (FHE) can impose a computational overhead of 10,000x to 1,000,000x compared to plaintext operations, making a standard database query that takes milliseconds on unencrypted data balloon to several minutes or even hours.

Secure Multi-Party Computation (MPC) takes a different approach by distributing the computation across multiple parties who each hold a share of the secret data. Instead of a single server performing heavy cryptographic work, two or more agency servers engage in a lightweight interactive protocol. This results in a vastly lower computational overhead, typically 10x to 100x slower than plaintext, and crucially, it avoids the ciphertext expansion problem that plagues HE. The trade-off is network communication: MPC protocols require multiple rounds of interaction, making them highly sensitive to network latency and bandwidth between the participating agencies.

The key trade-off: If your priority is a non-interactive, single-server architecture where data can be sent to an untrusted cloud for processing and returned later, choose Homomorphic Encryption. This model fits batch processing of large, non-urgent datasets where latency is measured in hours, not seconds. If you prioritize real-time, interactive queries where a citizen service agent needs an immediate response by querying a live database jointly held by the health and justice departments, choose Secure Multi-Party Computation. The latency of a few network round trips (typically 50-200ms in a dedicated inter-agency network) is far more practical than the multi-minute delay of an FHE query.

For the specific use case of real-time inter-agency AI inference—such as a law enforcement officer querying a combined social services and criminal justice database to assess risk during a live interaction—the latency requirements are non-negotiable. A 2024 pilot by the U.S. Defense Advanced Research Projects Agency (DARPA) on privacy-preserving analytics found that optimized 3-party MPC protocols could execute complex queries on distributed data in under 2 seconds, while equivalent FHE-based approaches required over 30 minutes. Therefore, MPC is the pragmatic choice for operational, time-sensitive government workflows.