TL;DR

HHEML is an end-to-end hardware-accelerated hybrid-homomorphic-encryption (HHE) framework targeting edge devices: a PYNQ-Z2 FPGA runs a pipelined Pasta symmetric cipher on the client, encrypted blocks are shipped over Ethernet to a server that transciphers them to FHE (via GuardML) and runs encrypted inference, cutting client-side latency from 356 ms to 6.5 ms on MNIST [§Abstract, §VI.B].

Problem and motivation

Pure FHE-based PPML imposes prohibitive computational and ciphertext-expansion overhead on resource-constrained IoT/edge clients [§I]. HHE mitigates this by using a lightweight, FHE-friendly symmetric cipher (Pasta) on the client and transciphering to FHE on the server [§I]. Prior HHE work is either software-only (high client latency) or hardware-only without integration into a complete PPML pipeline; HHEML targets that gap with a co-designed FPGA + Ethernet + FHE-server flow [§II.C]. Threat model: the computation server is untrusted (honest-but-curious is implied by the FHE-PPML setting); the client exclusively holds the secret key [§I].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
MNIST Image classification (encrypted inference benchmark) Not reported 28x28 grayscale images Encoded into 784-word plaintext blocks of 32-bit words [§VI.A]

Pipeline diagram

flowchart LR
    subgraph Client_FPGA [Client - PYNQ-Z2]
        P[Plaintext feature vec] --> PS1[PS orchestration]
        PS1 -. AXI-Stream/DMA .-> PL[Pasta-4 PL accelerator]
        PL --> SE[Pasta ciphertext]
    end
    subgraph Server [Server - GuardML]
        DEC[HHE Decomp transcipher] --> EVAL[HHE Eval on ML model]
    end
    subgraph Client_dec [Client - PYNQ-Z2]
        PL2[Pasta PL decrypt] --> OUT[Prediction]
    end
    SE -. Ethernet ct .-> DEC
    EVAL -. SE-wrapped result .-> PL2

Pipeline steps (text)

  1. PS collects a plaintext feature vector (e.g., an MNIST image as 784 32-bit words) [§IV.A, §VI.A].
  2. PS streams the data to the PL Pasta-4 accelerator over AXI4-Stream with DMA [§IV.B, §V.D].
  3. PL generates the SHAKE128 keystream with two parallel XOF modules and masks the plaintext in 17-word packets per round, producing a Pasta ciphertext [§V.A, §V.E].
  4. PS transmits the Pasta ciphertext plus the (SE-protected) symmetric key over Ethernet to the server [§IV.B].
  5. Server’s GuardML “HHE Decomp” transciphers Pasta ciphertext into FHE ciphertext [§IV.A, Fig. 2].
  6. Server runs homomorphic evaluation of the ML model on the FHE ciphertext [§IV.A].
  7. Server re-encodes the encrypted result as a symmetric ciphertext and returns it to the client [§IV.A].
  8. PL decrypts the returned ciphertext in hardware to produce the final prediction [§IV.A].

Architecture diagram

The paper does not specify the server-side ML model architecture; it reuses the GuardML inference stack [§IV.A, §VI.B]. The HHEML hardware datapath itself is shown below in lieu of an NN diagram.

flowchart LR
    I([Plaintext 17-word block]) --> XOF1[SHAKE128 XOF #1]
    I --> XOF2[SHAKE128 XOF #2]
    XOF1 --> MATGEN[MatGen]
    XOF2 --> MATGEN
    MATGEN --> MATMUL[MatMul affine A_j]
    MATMUL --> SUB[S'/S Sboxes]
    SUB --> ADD[Vec Add + round const]
    ADD --> MIX[Mix / final round]
    MIX --> O([Pasta ciphertext block])

Results

Metric This paper Baseline Hardware
1-round encryption 34.3 µs 66.1 µs (Pasta on Edge) PYNQ-Z2 @ 75 MHz [Table I]
MNIST encryption 1,553.4 µs 3,106.7 µs (Pasta on Edge) PYNQ-Z2 @ 75 MHz [Table I]
Rounds per MNIST image 24 47 (single-XOF) — [Table II]
Relative throughput 1.95x 1.00x (single-XOF) PYNQ-Z2 [Table II]
Client internal comm 6.5 ms — PYNQ-Z2 [Table III]
Client to server comm 0.4 ms 356 ms (GuardML SW) Ethernet LAN [Table III]
Total client-side latency 6.9 ms 356 ms (GuardML) — [Table III]
Server-side decomp (MNIST) 9,397.5 s 9,397.5 s (unchanged) Intel i7-10750H @ 2.6 GHz, 16 GB RAM [Table IV]
Server-side eval (MNIST) 54.9 s 54.9 s (unchanged) Intel i7-10750H @ 2.6 GHz, 16 GB RAM [Table IV]
Power (client) 1.505 W 1.2 W (Pasta on Edge) PYNQ-Z2 [Table I]
LUTs / FFs / DSP / BRAM 34,419 / 18,372 / 64 / 2 23,736 / 11,132 / 64 / 0 PYNQ-Z2 [Table I]
Headline accuracy Not reported Not reported —

Note: per the template rules, the paper only reports server-side end-to-end transciphering + evaluation totals (9,397.5 s + 54.9 s for MNIST batch processing) without a per-image FHE inference number, so single_inference_seconds: N/A.

Limitations and assumptions

Code and artifacts

Not released (no repository URL provided in the paper).

Open questions