Fully Homomorphic Encryption: Origins, Development, Applications, and Engineering Challenges in Privacy Computing

01 Current Non‑Fully Homomorphic Privacy Computing Defects

Typical privacy‑computing models involve multiple data owners who wish to jointly compute statistics or machine‑learning models without revealing raw data. Existing techniques such as semi‑homomorphic encryption and secret sharing suffer from non‑generic computation (one protocol per use case), exposure of computation logic, and high communication overhead due to many interaction rounds.

These issues are illustrated with logistic‑regression protocols that require all parties to follow the same detailed protocol, making updates costly and exposing proprietary algorithms. Multi‑round interactions also strain network bandwidth and stability, especially in financial environments.

02 Fully Homomorphic Encryption Advantages

Fully homomorphic encryption (FHE) allows unlimited addition and multiplication on ciphertexts, enabling arbitrary function evaluation without decrypting data. This eliminates the need for shared protocols, reduces interaction rounds, protects computation logic, and eases network requirements.

03 Development History of Fully Homomorphic Encryption

The cryptographic lineage starts with RSA (1977), followed by randomized schemes like ElGamal (1985) and Paillier (1999). The quest for both additive and multiplicative homomorphism led to BGN (2005) and later generations of FHE algorithms (first‑gen 2009, subsequent faster versions, and CKKS for floating‑point support in 2017).

Parallel engineering progress produced open‑source libraries such as IBM HElib (BGV), Microsoft SEAL (BFV/CKKS), and hardware acceleration initiatives by the U.S. Department of Defense, Intel, and Google.

04 Standardization and Algorithm Selection

Choosing a secure algorithm requires considering long‑term resistance (e.g., post‑quantum security) and compliance with emerging standards from NIST, academia, and industry.

05 Application Scenarios

Outsourced computing: encrypt data, send to cloud, obtain results without data leakage.

Model training: encrypted labels enable collaborative training with minimal communication.

Joint risk control: banks and data providers keep their data and rules confidential while jointly evaluating credit decisions.

Financial engineering: encrypted client holdings are processed by proprietary models without exposing either side’s secrets.

06 Engineering Practice

Open‑source libraries (HElib, SEAL, etc.) provide building blocks. Protocol selection depends on data type: BFV/BGV for integer vectors, CKKS for floating‑point vectors, and FHEW/TFHE for piecewise or non‑continuous functions.

Operator libraries implement logical gates (AND, OR, NOT) and arithmetic on ciphertexts. Long computation chains accumulate error; ciphertext refresh (bootstrapping) is used to reset noise after a bounded number of operations.

Tencent Angel PowerFL integrates CKKS, offering fast homomorphic comparison and federated logistic/linear regression, achieving several‑fold speedups over Paillier‑based solutions.

Overall, fully homomorphic encryption offers a promising path to secure, scalable privacy computing, especially for finance and machine‑learning workloads where data confidentiality and algorithm secrecy are paramount.