TL;DR

A Harvard undergraduate honors thesis (March 20, 2023) that gives a from-first-principles exposition of FHE, covering LWE/RLWE, the GSW and CKKS schemes, and a survey of FHE-based privacy-preserving machine learning. Its single technical contribution is a modified xMK-CKKS-based “SecureFederatedAveraging” protocol that supports weighted averaging and client subsampling [§4.3.3].

Problem and motivation

Machine learning services on personal data conflict with privacy demands and regulation (HIPAA, GDPR), and patient reluctance further limits genetic studies [§1, p. 5; §4.0.1, p. 49]. FHE resolves this by enabling computation on ciphertexts. The thesis targets a “mathematically mature student with no background in cryptography, number theory, and/or machine learning” and aims to take such a reader to the research frontier [§1.1, p. 6]. The threat model assumed throughout for ML applications is Honest-but-Curious servers [§4, p. 49].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
MNIST Handwritten-digit classification Not reported 28x28 grayscale images Mentioned only as the benchmark for CryptoNets/SHE; not used directly in the thesis [§4.2.1, p. 55]
Tumor genetic data Tumor classification Not reported Genetic Referenced from [HPC+22]; “millions of genetic datapoints”, limited to CKKS ciphertexts of 65,536 elements [§4.2, p. 54]
Chest X-rays / COVID-19 Classification Not reported Medical imaging Referenced from [WCK+22] as motivation for federated learning [§4.3, p. 56]

Pipeline diagram

flowchart LR
    subgraph ClientK [Client k]
        A[Local SGD on D_k] --> B[Scale by D_k]
        B --> C[Encrypt w_scaled<br/>under b_aggr]
    end
    subgraph Server
        D[Aggregate ciphertexts<br/>EVALADD] --> E[Request decryption<br/>shares]
    end
    subgraph ClientDec [Client k]
        F[Compute D_k share] --> G[Server runs SUMDEC]
        G --> H[Divide by n<br/>= weighted avg]
    end
    C -. ciphertext .-> D
    E -. request .-> F
    F -. dec share .-> G

Pipeline steps (text)

  1. Setup phase: server samples public parameter a; each client runs CKKS key-gen to produce (s_i, b_i); server publishes aggregate public key b_aggr = Σ b_i [Def. 4.3.2, p. 58].
  2. In round t, server selects random subset S_t of clients (size at least 2 to prevent single-client leakage) [§4.3.3, p. 60].
  3. Each chosen client runs batched SGD on its local data D_k for E epochs and scales the resulting weights by |D_k| [Alg. 2, p. 60].
  4. Client encrypts scaled weights under b_aggr and sends ciphertext to server.
  5. Server homomorphically adds the ciphertexts via EVALADD to obtain c_+.
  6. Each client computes a partial decryption share D_i = s_i · c_1 + e*_i and sends to server.
  7. Server runs SUMDEC(c_+, D_1, ..., D_N) and divides by n = Σ|D_k| to obtain weighted-average global weights [Alg. 2, p. 60].
  8. Server distributes new weights w_{t+1} to clients and repeats.

Architecture diagram

This is a thesis/survey, so the “architecture” is the chapter taxonomy of the document rather than a neural network.

flowchart LR
    I([Ch.1 Intro & Crypto Prelims]) --> L[Ch.2 LWE / RLWE]
    L --> G[Ch.3.3 GSW Scheme]
    L --> K[Ch.3.4 CKKS Scheme]
    G --> B[Bootstrapping]
    K --> H[Ch.4 FHE for ML]
    H --> S[Survey: CryptoNets, SHE, CraterLake]
    H --> M[Multikey FHE / xMK-CKKS]
    M --> F[SecureFederatedAveraging]
    F --> C([Ch.5 Conclusion])

Results

The thesis reports no original empirical numbers. It cites results from surveyed works:

Metric This paper Baseline Hardware
Original empirical results None reported — —
CryptoNets accuracy (cited) 98.95% on MNIST, >50,000 images/hour Plaintext SOTA at the time: 99.79% [WZZ+13] Not reported [§4.2.1, p. 55]
[CdWM+17] accuracy (cited) 99.3% on MNIST CryptoNets 98.95% Not reported [§4.2.1, p. 55]
SHE accuracy (cited) 99.77% on MNIST in half CryptoNets’ time; shallower variant 99.54% in <4% of CryptoNets time CryptoNets Not reported [§4.2.1, p. 55]
CraterLake (cited) ~10x speedup over CPU on CKKS ops Standard CPU CKKS Hardware accelerator [§4.2.1, p. 55]

Limitations and assumptions

Surveys and positions against, but does not benchmark experimentally:

Code and artifacts

Not released. No implementation is described in the thesis.

Extra diagrams (optional)

Threat model

flowchart LR
    C1[Client 1<br/>data D_1, key s_1] -- enc(w_1) --> S[Server<br/>Honest-but-Curious]
    C2[Client 2<br/>data D_2, key s_2] -- enc(w_2) --> S
    CN[Client N<br/>data D_N, key s_N] -- enc(w_N) --> S
    S -- request --> C1
    S -- request --> C2
    S -- request --> CN
    C1 -- dec share D_1 --> S
    C2 -- dec share D_2 --> S
    CN -- dec share D_N --> S
    S -- w_t+1 --> C1
    S -- w_t+1 --> C2
    S -- w_t+1 --> CN

The server is Honest-but-Curious; clients do not reveal data or local weights to each other or the server [§4, p. 49; §4.3.3, p. 60–61].

Federated round

sequenceDiagram
    participant C as Clients in S_t
    participant A as Server
    A->>C: send current w_t
    C->>C: local SGD, scale by |D_k|
    C->>A: enc_{b_aggr}(|D_k| * w_k)
    A->>A: EVALADD over ciphertexts
    A->>C: request decryption shares
    C->>A: D_i = s_i * c_1 + e*_i
    A->>A: SUMDEC and divide by n
    A->>C: distribute w_{t+1}

One round of SecureFederatedAveraging from Algorithm 2 [§4.3.3, p. 60].

Open questions