Academic Paper Alerts — New FHE Research Relevant to Zama's Roadmap

# fhe-paper-monitor.yaml
name: fhe-academic-watcher
schedule: "0 9 * * 1"  # Every Monday 9am UTC
skills:
  - deep-research
  - hacker-news-scraper  # for community discussion

sources:
  arxiv:
    categories: ["cs.CR", "cs.CC", "cs.LG", "cs.DS"]
    search_terms:
      primary:
        - "fully homomorphic encryption"
        - "TFHE"
        - "homomorphic encryption"
        - "FHE"
      secondary:
        - "lattice-based cryptography"
        - "bootstrapping" AND "encryption"
        - "privacy-preserving computation"
        - "encrypted machine learning"
        - "confidential computing"
    lookback_days: 7

  iacr_eprint:
    search_terms: ["FHE", "TFHE", "homomorphic", "bootstrapping"]
    lookback_days: 7

  conferences:
    track:
      - "Crypto 2026"
      - "Eurocrypt 2026"
      - "CHES 2026"
      - "IEEE S&P 2026"
      - "NeurIPS 2026"  # FHE+ML intersection
    check: "accepted papers list when available"

relevance_scoring:
  high:
    - "TFHE" in title or abstract
    - "bootstrapping" improvement with benchmarks
    - direct comparison to Zama's libraries (concrete, tfhe-rs)
    - "programmable bootstrapping" or "PBS"
  medium:
    - general FHE scheme improvements
    - FHE compiler or tooling papers
    - encrypted ML inference or training
    - lattice parameter optimization
  low:
    - theoretical FHE (no implementation/benchmarks)
    - tangentially related crypto papers
    - survey papers (unless comprehensive)

  filter: "Include HIGH and MEDIUM. Exclude LOW unless
           author is known FHE researcher."

output:
  weekly_digest: workspace/research/fhe-papers/weekly/
  format: |
    For each paper:
    - Title + authors + link
    - One-paragraph plain-English summary
    - Relevance to Zama's roadmap (specific connection)
    - Key result or claim
    - Methodology strength (peer-reviewed? benchmarked? reproducible?)
    - Recommended action: READ / SKIM / CITE / IGNORE

## FHE Research Digest — Week of 2026-03-09
**Papers scanned:** 54
**Papers included:** 3 HIGH relevance, 2 MEDIUM

---

### HIGH RELEVANCE

**1. "Faster Bootstrapping for TFHE via Optimized NTT on GPU"**
Authors: Chen et al. (EPFL)
Link: arxiv.org/abs/2603.XXXXX
**Summary:** Achieves 2.8x speedup on TFHE bootstrapping by
implementing a GPU-optimized Number Theoretic Transform. Benchmarked
on NVIDIA A100, directly compared against tfhe-rs performance.
**Zama relevance:** DIRECT — this optimization could be integrated
into Zama's GPU backend. The NTT optimization is compatible with
Zama's existing architecture. Contact authors about collaboration.
**Key claim:** 12ms bootstrapping (vs 34ms baseline tfhe-rs)
**Methodology:** Strong — benchmarks reproduced, code available
**Action:** READ — forward to Zama's GPU team immediately

**2. "Programmable Bootstrapping Without Noise Growth: A New Approach"**
Authors: Ducas, Stehlé (CWI Amsterdam, ENS Lyon)
Link: iacr.org/eprint/2026/XXX
**Summary:** Proposes a modified PBS scheme that eliminates noise
accumulation during evaluation, theoretically enabling unlimited
circuit depth without bootstrapping refresh.
**Zama relevance:** CRITICAL — if the claims hold, this could
fundamentally change TFHE circuit design. Eliminates the key
performance bottleneck in deep FHE circuits.
**Key claim:** Unlimited depth PBS with O(1) noise growth
**Methodology:** Theoretical — no implementation yet. Needs
verification by Zama's crypto team.
**Action:** READ — schedule review session with Zama crypto team

**3. "TFHE-Based Private Inference for Transformer Models:
Practical Benchmarks"**
Authors: Park et al. (Seoul National University)
Link: arxiv.org/abs/2603.YYYYY
**Summary:** First practical benchmarks for running a 125M parameter
transformer model under TFHE encryption. End-to-end inference in
47 seconds on standard hardware.
**Zama relevance:** HIGH — validates the FHE+ML direction. Their
approach uses techniques similar to Zama's Concrete ML but with
different encoding strategies. Potential improvement path.
**Key claim:** 47-second encrypted inference (125M params)
**Methodology:** Good — code available, benchmarks reproducible
**Action:** READ — compare encoding strategy with Concrete ML

---

### MEDIUM RELEVANCE

**4. "A Survey of FHE Compiler Optimizations (2023-2026)"**
Survey paper covering recent compiler advances. Useful reference
but no new results.
**Action:** SKIM — save as reference

**5. "Lattice Parameter Selection for Multi-Key TFHE"**
Parameter optimization for MK-TFHE schemes. Relevant if Zama
expands to multi-key applications.
**Action:** SKIM — relevant for future roadmap items

mrchief deep-research \
  --query "Analyze arxiv.org/abs/2603.XXXXX — implications for
  Zama's TFHE performance benchmarks, integration feasibility
  with tfhe-rs, and comparison with existing GPU optimizations" \
  --depth deep \
  --output workspace/research/fhe-papers/deep-dives/chen-gpu-ntt.md