Privacy-Enhancing Technologies: How FHE, MPC

The Privacy-Utility Tradeoff Is Being Solved

For decades, data privacy and data utility were treated as fundamentally opposed. To analyze data, you had to access it in the clear. To protect privacy, you had to restrict access, often rendering the data useless for the analytics, machine learning, and collaborative computation that drive modern business and research. This tradeoff shaped regulations (GDPR’s data minimization principle), architectures (centralized data lakes with access controls), and entire business models (data brokers extracting value from aggregated personal data).

Privacy-Enhancing Technologies (PETs) are dissolving this tradeoff. Three technologies — Fully Homomorphic Encryption (FHE), Multi-Party Computation (MPC), and differential privacy — now enable organizations to compute on data without ever seeing it in plaintext. Gartner projected that by 2025, 60% of large organizations would use at least one privacy-enhancing computation technique for analytics, AI, or cloud computing. Actual adoption has run slightly behind that forecast but is accelerating sharply: Grand View Research valued the global PET market at $3.1 billion in 2024, and multiple analyst firms project it will exceed $12 billion by 2030.

The shift is driven by converging pressures: GDPR and its global successors imposing strict limits on data sharing, the cumulative EUR 5.88 billion in GDPR fines levied through early 2025 creating real financial consequences, the healthcare and financial sectors needing cross-institutional analytics without exposing patient or customer records, and the AI industry requiring massive training datasets while facing growing restrictions on data collection. PETs offer a technical resolution to what was previously a legal and ethical impasse.

Fully Homomorphic Encryption: Computing on Ciphertext

Fully Homomorphic Encryption is arguably the most transformative PET — and the one that has taken the longest to become practical. First theorized by Rivest, Adleman, and Dertouzos in their 1978 paper “On Data Banks and Privacy Homomorphisms” and first constructed by Craig Gentry in his landmark 2009 Stanford PhD thesis, FHE allows arbitrary computations on encrypted data. You encrypt your data, send the ciphertext to a server, the server performs computations on the ciphertext, and returns encrypted results that you decrypt — the server never sees the plaintext at any point.

The historical barrier was performance. Gentry’s original scheme required roughly 30 minutes per single logic gate operation — many orders of magnitude slower than computing on plaintext. Fifteen years of cryptographic research have reduced this overhead dramatically. Current FHE schemes (BGV, BFV, CKKS, TFHE) achieve overheads of 1,000x to 100,000x depending on the computation type — still significant but now practical for specific workloads. Zama, the Paris-based startup that raised $73 million in Series A funding in March 2024 and a further $57 million in 2025 to become the first FHE unicorn, has emerged as the leading FHE tooling company. Its open-source library Concrete enables developers to write FHE applications in Python, and its Concrete ML product allows machine learning inference on encrypted data with performance overhead of 10-100x for common models.

Real-world deployments are emerging. CryptoNets, Microsoft’s research project, demonstrated that a neural network can classify encrypted MNIST images with 99% accuracy at a throughput of over 51,000 predictions per hour. In healthcare, Duality Technologies — which has raised over $50 million in venture funding and partnered with organizations including Mastercard on privacy-preserving data collaboration — deploys FHE-based analytics enabling hospitals to collaborate on patient data analysis without sharing records. Swiss Post uses homomorphic encryption in its e-voting system deployed across Swiss cantons, ensuring vote confidentiality during tallying. Intel’s FHE acceleration efforts — including the HEXL software library that leverages AVX-512 instructions on Xeon processors, and a custom ASIC being developed under the DARPA DPRIVE program — signal that dedicated hardware support will further reduce the performance gap within the next hardware generation.

MPC and Differential Privacy: Complementary Approaches

Multi-Party Computation (MPC) solves a different but related problem: how can multiple parties jointly compute a function over their combined data without any party revealing its individual input to the others? First formalized by Andrew Yao in his 1982 paper “Protocols for Secure Computations” — which introduced the famous Millionaires’ Problem — and later concretized with the Garbled Circuits protocol in 1986, MPC has evolved from a theoretical curiosity into a production technology. The key insight is that data is split into “shares” distributed across multiple servers — no single server holds enough information to reconstruct any individual record, but the servers can collaboratively compute results that are mathematically equivalent to computing on the combined dataset.

MPC’s most visible production deployment is in the financial sector. The Danish company Partisia (which spun out Partisia Blockchain, raising over $54 million including token sales) uses MPC to enable financial institutions to perform joint anti-money laundering (AML) analysis across bank boundaries — each bank contributes transaction patterns without revealing customer data to other banks, enabling detection of money laundering schemes that no single bank could identify alone. In Denmark, a January 2008 auction for sugar beet production contracts used MPC to determine market-clearing prices across more than 1,200 bidders without any farmer revealing their bid to competitors — one of the earliest and most cited real-world MPC deployments. More recently, major cryptocurrency custodians including Fireblocks ($8 billion valuation, over $10 trillion in transactions secured) use MPC to protect private keys, distributing key shares across multiple servers so that no single compromised server can sign transactions.

Differential privacy takes yet another approach: instead of encrypting data, it adds carefully calibrated mathematical noise to query results, guaranteeing that the inclusion or exclusion of any individual’s data cannot be detected from the output. Apple pioneered large-scale differential privacy deployment in 2016, using it to collect usage statistics from hundreds of millions of iOS devices — including popular emojis, new words, and Safari browsing patterns — without being able to identify individual users’ behavior. Google’s RAPPOR (Randomized Aggregatable Privacy-Preserving Ordinal Response), released in 2014, applies differential privacy to Chrome usage data. The U.S. Census Bureau used differential privacy for the 2020 Census — a controversial decision that sparked debate about the tradeoff between privacy guarantees and statistical accuracy for small geographic areas. The privacy budget (epsilon parameter) determines how much noise is added: lower epsilon means stronger privacy but noisier results, creating a tunable knob that organizations must calibrate to their specific risk tolerance.

Performance, Adoption Barriers, and the Road Ahead

Despite dramatic progress, PETs face real adoption barriers that separate current reality from the theoretical promise. Performance overhead remains the primary constraint for FHE: while 10-100x overhead is acceptable for batch analytics, it makes real-time applications (fraud detection in milliseconds, interactive queries) impractical. MPC introduces communication overhead — the servers performing joint computation must exchange messages proportional to the circuit complexity, making wide-area deployments latency-sensitive. Differential privacy’s noise injection can degrade utility for small datasets or rare events, limiting its applicability where precision matters at the individual level.

The tooling ecosystem is maturing but remains immature compared to conventional data infrastructure. FHE programming requires understanding noise budgets, bootstrapping operations, and plaintext slot management — concepts alien to most software engineers. Zama’s Concrete and Microsoft’s SEAL library have simplified FHE development significantly, but the developer experience remains closer to assembly language than Python in terms of abstraction level. MPC frameworks like MP-SPDZ (University of Bristol) and ABY (TU Darmstadt) require cryptographic expertise to configure correctly. The talent pool of engineers who can design and implement PET-based systems is measured in the low thousands globally, creating a bottleneck for enterprise adoption.

The convergence trajectory is clear despite these barriers. Hardware acceleration (Intel’s HEXL library and DARPA DPRIVE ASIC, GPU-accelerated MPC), improved compiler technology (automatically converting standard programs into FHE or MPC circuits), and standardization efforts (the HomomorphicEncryption.org consortium, ISO/IEC 27559 for differential privacy) are collectively reducing the barrier to adoption. Zama’s trajectory — $73 million Series A in 2024 followed by unicorn status in 2025 — signals venture capital’s conviction that FHE will become a mainstream infrastructure technology. Within five years, PETs are likely to be embedded in cloud platforms as standard offerings — AWS, Azure, and GCP all have active PET research programs — making privacy-preserving computation a configuration choice rather than a specialized engineering project.

Frequently Asked Questions

Sources & Further Reading

• Zama Series A — $73M for Fully Homomorphic Encryption (TechCrunch, 2024)
• Zama Becomes First FHE Unicorn with $57M Raise (CoinDesk, 2025)
• Grand View Research — Privacy Enhancing Technologies Market Report
• GDPR Enforcement Tracker — Cumulative Fines Data
• CMS GDPR Enforcement Tracker Report 2024/2025
• Craig Gentry — Fully Homomorphic Encryption PhD Thesis (Stanford, 2009)
• Microsoft CryptoNets — Neural Networks on Encrypted Data
• Danish Sugar Beet Auction — First Large-Scale MPC Deployment (Wikipedia)
• Partisia — World’s First Commercial MPC Deployment
• Fireblocks — MPC-Based Digital Asset Security ($8B Valuation)
• Apple Differential Privacy Technical Overview
• Google RAPPOR — Privacy-Preserving Reporting Algorithms
• Intel HEXL — Homomorphic Encryption Acceleration Library
• Swiss Post E-Voting Security and Cryptography
• HomomorphicEncryption.org — Industry Standardization Consortium
• Microsoft SEAL — Homomorphic Encryption Library
• U.S. Census Bureau — Differential Privacy and the 2020 Census
• Duality Technologies — Privacy-Preserving Data Collaboration