TL;DR

Minelli’s PhD thesis develops FHE techniques motivated by ML applications: a TFHE-based framework (FHE-DiNN) for evaluating discretized neural networks with cost independent of network depth, a Gaussian-leftover-hash-lemma technique for circuit privacy in GSW without noise flooding, and a homomorphic-encryption PIR protocol for cross-border law-enforcement queries [Abstract; §1.4].

Problem and motivation

The thesis addresses three privacy gaps in the user-server outsourcing scenario [§1.2]:

Threat model is honest-but-curious throughout (the thesis explicitly leaves malicious adversaries as an open question) [§7.2.2, Question 7.6].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
MNIST 10-class digit classification 60,000 / 10,000 (standard split implied by full test-set evaluation) [§4.5.5.3, Table 4.5] 28x28 grayscale images, binarized to {-1, +1} at threshold 128 Input shape 784, single-bit per pixel after binarization [§4.5.5.1]
Law-enforcement records (synthetic) PIR / OT N = 10,000 records [§6.4.3] Tabular (ID + features such as DNA, fingerprints, car plate) Encrypted inverted-index database with PRF-tokenized feature keys and DGHV-encrypted ID polynomials [§6.2]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Binarize image] --> B[Pack 784 pixels into 1 TLWE]
    end
    subgraph Server
        C[Mul by weight poly] --> D[Extract LWE]
        D --> E[KeySwitch N to n]
        E --> F[BlindRotate + sign]
        F --> G[Layer-2 multisum]
    end
    subgraph Client_dec [Client]
        H[Decrypt 10 scores] --> I[argmax label]
    end
    B -. ciphertext .-> C
    G -. 10 ciphertexts .-> H

Pipeline steps (text)

  1. Client binarizes the MNIST image to {-1, +1} pixels [§4.5.5.1].
  2. Client packs all 784 pixels into one TLWE ciphertext encoding the polynomial Σ_i x_i X^i [§4.5.4.1].
  3. Server multiplies the TLWE ciphertext by the precomputed FFT representation of the weight polynomial Σ_i w_i X^{-i} for each hidden neuron, producing 30 (or 100) TLWE ciphertexts [§4.5.5.3, step 2].
  4. Server runs Extract on each to recover a 1024-LWE ciphertext encoding the multisum [step 3].
  5. Server applies KeySwitch to drop the dimension from N = 1024 to n = 450 [step 4].
  6. Server runs the sign-bootstrapping: BlindRotate with the testVector encoding sign, then Extract — this both refreshes noise and applies sign(·) for free [§4.5.3.2, step 5].
  7. Server takes the 30 (or 100) refreshed ciphertexts, applies the second-layer multisum with the output weights, yielding 10 N-LWE ciphertexts (one per class) [step 6].
  8. Server returns the 10 ciphertexts; client decrypts and takes argmax [step 7].

Architecture diagram

This thesis spans three contributions; a single network-layer diagram does not capture the whole work. We provide one diagram for FHE-DiNN’s classifier and a second for the thesis chapter structure.

FHE-DiNN classifier (784:H:10, H in {30, 100})

flowchart LR
    I([Input 784 binary]) --> FC1[FC H neurons]
    FC1 --> A1[sign via TFHE bootstrap]
    A1 --> FC2[FC 10 neurons]
    FC2 --> O([10 scores -> argmax])

Thesis contribution structure

flowchart LR
    P[Preliminaries Ch.2-3] --> N[Ch.4 FHE-DiNN TFHE for NNs]
    P --> C[Ch.5 Circuit privacy for GSW]
    P --> R[Ch.6 PIR via extended DGHV]
    N --> X[Ch.7 Conclusions / open questions]
    C --> X
    R --> X

Results

Metric This paper Baseline Hardware
Plaintext accuracy, FCNN-30 (DiNN + sign) 93.55% [Table 4.2] 94.76% original real-valued NN N/A
Plaintext accuracy, FCNN-100 (DiNN + sign) 96.43% [Table 4.2] 96.75% original real-valued NN N/A
Homomorphic accuracy, FHE-DiNN 30 93.71% (273 disagreements vs. clear, of 10,000) [Table 4.5] Cryptonets 98.95% (deeper net) [Table 4.6] Intel Core i7-4720HQ @ 2.60GHz, 1 core
Homomorphic accuracy, FHE-DiNN 100 96.35% (127 disagreements) [Table 4.5] Cryptonets 98.95% Same
Single-image inference time, FHE-DiNN 30 (orig BlindRotate) 0.515 s [Table 4.5] Cryptonets 570 s per batch of 8192 (= 0.07 s amortized) [Table 4.6] Same
Single-image inference time, FHE-DiNN 30 (unfolded BlindRotate) 0.491 s [Table 4.5] — Same
Single-image inference time, FHE-DiNN 100 1.65–1.68 s [Table 4.5] — Same
Single bootstrapping ≈15 ms [§4.5.5] — Same
Ciphertext size per image ≈8.2 kB [§4.5.5, Table 4.6] Cryptonets 586 MB unbatched / 73.3 kB amortized [Table 4.6] —
PIR simple-query bandwidth ≈1510·d KB (d = polynomial degree) [Table 6.3] — DGHV-based prototype, Python/Flask
PIR conjunctive-query bandwidth ≈6795·(d1+d2) KB [Table 6.3] — Same

Limitations and assumptions

Code and artifacts

Not released as a single repository in the thesis itself. The FHE-DiNN implementation extends the public TFHE library [CGGI16a] available on GitHub [§4.5.5]; the PIR prototype is a custom Python/Flask implementation, no release URL stated [§6.4.3, ref. [10]]. License: Not reported.

Extra diagrams (optional)

Threat model (PIR protocol)

flowchart LR
    FR[France querying party] -- search token --> DE[Germany DB server]
    FR -- query plus token plus evidence --> JU[Judge trusted party]
    JU -- signed enc poly --> FR
    FR -- signed enc poly --> DE
    DE -- record ids --> FR

Circuit-privacy core step (Ch.5)

flowchart LR
    V0[Enc v0] --> S[V_out = C * G^-1_rand V1 + G - C * G^-1_rand V0 + Gaussian shift]
    V1[Enc v1] --> S
    C[Enc x] --> S
    S --> V[Enc v_x circuit-private]

Open questions