Fully homomorphic encryption and the dawn of a private internet
Fully homomorphic encryption (FHE) promises the ability to run computations on encrypted data so that cloud providers never see plaintext, raising hopes for truly private search, AI, and database services. Commenters largely agree the math is sound but stress that current schemes are orders of magnitude slower and larger than plaintext, making most real-world use cases — like web-scale search or LLM inference — economically and technically impractical for now. Many expect FHE to remain niche, used mainly in high‑value, regulated, or narrowly scoped scenarios alongside alternatives like confidential computing and specialized searchable encryption, with broader adoption constrained more by business incentives and user apathy than by cryptography alone.
Performance and scalability limits
- Many commenters argue current and foreseeable FHE is orders of magnitude slower (often cited ~1000×, possibly far worse) than plaintext due to bootstrapping and huge ciphertexts.
- Even without bootstrapping, ciphertext blowup (often ~10³× larger) implies massive extra memory bandwidth and compute that hardware advances alone are unlikely to erase.
- Latency impacts are framed as unacceptable for most user-facing tasks (e.g., milliseconds → seconds/minutes; 30s → hours).
Search, databases, and index problems
- Fully private search over large indexes is highlighted as especially hard:
- Naively, encrypted-key search is O(n) instead of O(log n); recent PIR-style schemes with polylog(n) queries require extreme preprocessing and storage blowups (petabytes for modest DBs).
- For “FHE Google,” either the server must encrypt huge indexes per client or do near-linear work per query, both viewed as impractical.
- Some systems mix FHE/PIR with partial leakage (e.g., hashed prefixes, subsets, anonymous networks), trading strict privacy for performance.
Use cases and economics
- Strong consensus: generic “private internet” via FHE is not economically viable soon. A privacy‑first Google‑scale service would be vastly more expensive and far slower; few users would pay enough.
- People often prefer free, data‑harvesting services; privacy‑preserving alternatives already exist but remain niche.
- FHE is seen as promising for narrow, high‑value, low‑throughput tasks: finance, regulated data sharing, some medical or government/military computations, “data owner vs model owner” scenarios.
Alternatives to FHE
- Confidential computing (TEEs: SGX/TDX/SEV/ARM TEE) is repeatedly described as the only realistic way to do private LLM inference and many cloud workloads, despite hardware‑trust issues and past breaks.
- Specialized searchable encryption and PIR schemes can give near‑plaintext search performance for specific patterns, with FHE reserved for small filtered subsets.
- Self‑hosting and encrypted backups are presented as a simpler, cheaper privacy route for many personal data uses.
How FHE works & security nuances
- Multiple explanations outline homomorphism: operations on ciphertext correspond to operations on plaintext, enabling arbitrary circuits via additions/multiplications.
- Some struggle with the intuition that “being able to compute means weaker encryption”; others note:
- FHE is malleable (not NM‑CCA2) but can still be IND‑CPA/semantically secure.
- Computation may leak allowed structure (e.g., which operations were run) but not plaintext, and “circuit privacy” research aims to hide even that.
Privacy, incentives, and politics
- Commenters doubt big providers will voluntarily adopt FHE that blocks data harvesting without strong regulation or market pressure.
- Governments may resist or demand backdoors; export controls and surveillance incentives are seen as major non‑technical obstacles.