ECHO: Efficient Covertly-Secure Three-party Computation with Applications to Private Machine Learning

This repository contains the experimental code for the paper ''ECHO: Efficient Covertly-Secure Three-party Computation with Applications to Private Machine Learning.'' It provides implementations and benchmarking harnesses for the protocols evaluated in the paper across arithmetic and binary settings, optimized for different network and end-to-end objectives.

Repository Contents (Protocols)

• Semi-honest protocols:

  • ABY3
  • ASTRA

• Covert protocols:

  • Covert MAC (Arithmetic, optimized for end-to-end)
  • Covert MAC (Arithmetic, optimized for online)
  • Covert DZKP (Arithmetic, optimized for WAN)
  • Covert ABFL (Binary, optimized for end-to-end)
  • Covert FLNW (Binary, optimized for online)
  • Covert DZKP (Binary, optimized for WAN)

• Malicious protocols:

  • SpdzWise Ring (Arithmetic, E2E optimal)
  • FLNW17+SPDZ2k (Arithmetic, Online optimal)
  • LEDH+24 (Arithmetic, WAN optimal)
  • FLNW17 (Binary, E2E optimal)
  • ABFL+17 (Binary, Online optimal)
  • LDHH+23 (Binary, WAN optimal)

Build and Run

1) Environment setup

First, set up the MP-SPDZ environment as described in README-MP_SPDZ.md.

Note: when running WAN benchmarks, you can use the traffic-control.sh script to simulate the network condition.

# Add a new network condition
./traffic-control.sh add <bandwidth> <latency>

# Remove a network condition
./traffic-control.sh clear

2) Build benchmark binaries

Use make to compile the benchmark executables:

make mul-benchmark.x and-benchmark.x and-benchmark2.x -j

make replicated-ring-party.x sy-rep-ring-party.x echo-prep-party.x echo-party.x covert-mac-party.x astra-prep-party.x astra-party.x -j

3) Run benchmarks

One command to run all NN benchmarks

Use the Python runner to orchestrate experiments and collect logs/averages with one command.

Single machine (spawns all required parties per experiment):

python3 run-NN.py                        # run all default experiments

Common flags:

• --round, --num_gates, --batch_size, --binary_batch_size: override global parameters
• --host, --port: network host/port passed into the benchmark binaries
• --repeats: number of runs per experiment (default: 3)
• --outdir: output directory (default: logs/<timestamp>)

Separate commands to run each microbenchmarks

The commands below illustrate how to benchmark each protocol family. Replace the environment variables as needed:

• ROUNDS: number of protocol iterations/rounds
• HOST: hostname or IP of the party 0 process
• PORT_BASE: base port number (ports will be derived per party)
• BATCH_SIZE: arithmetic batch size
• BINARY_BATCH_SIZE: binary batch size

Run each command with party id 0/1/2 in separate terminals or machines as appropriate.

Semi-honest protocols

• ABY3

./mul-benchmark.x 0 semi -r $ROUNDS -h $HOST -pn $PORT_BASE

• ASTRA

./mul-benchmark.x 0 astra -r $ROUNDS -h $HOST -pn $PORT_BASE (on Machine 0, 1, 2)

./mul-benchmark.x 0 astra-on -r $ROUNDS -h $HOST -pn $PORT_BASE (on Machine 1, 2. Where machine 1 uses party id=0 and machine 2 uses party id=1)

Covert protocols

• Covert MAC (Arithmetic, E2E)

./mul-benchmark.x 0 spdz -r $ROUNDS -h $HOST -pn $PORT_BASE -S 12 -b $BATCH_SIZE

• Covert MAC (Arithmetic, Online)

./mul-benchmark.x 0 offline -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE -q 16 -lm 2
./mul-benchmark.x 0 online  -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE

• Covert DZKP (Arithmetic, WAN)

./mul-benchmark.x 0 cov-dzkp -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE -q 16 -lm 2

This protocol automatically uses a 12-bit CCS ring gap.

• Covert ABFL (Binary, E2E)

./and-benchmark.x 0 cov-trip -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

• Covert FLNW (Binary, Online)

./and-benchmark2.x 0 offline -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE
./and-benchmark2.x 0 online  -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

• Covert DZKP (Binary, WAN)

./and-benchmark.x 0 cov-dzkp -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

Malicious protocols

• SpdzWise Ring (Arithmetic, E2E optimal)

./mul-benchmark.x 0 spdz -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE

• FLNW17 (Arithmetic, Online optimal)

./mul-benchmark.x 0 mal -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE

• LEDH+24 (Arithmetic, WAN optimal)

./mul-benchmark.x 0 ccs -r $ROUNDS -h $HOST -pn $PORT_BASE -b $BATCH_SIZE -lm 40 -q 16

This protocol automatically uses a 64-bit CCS ring gap.

• FLNW17 (Binary, E2E optimal)

./and-benchmark.x 0 mal -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

• ABFL+17 (Binary, Online optimal)

./and-benchmark.x 0 abfl -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

• LDHH+23 (Binary, WAN optimal)

./and-benchmark.x 0 uss -r $ROUNDS -h $HOST -pn $PORT_BASE -bb $BINARY_BATCH_SIZE

Notes

• The 0 in the example commands denotes the party id (0/1/2). Launch one instance per party.
• Some protocols require parameter tuning (-S, -q, -lm) and small source edits as noted above in Protocols/Ccs24MC.h.
• Ensure networking between parties is configured (firewall, NAT traversal) according to your deployment.