CO-Optimized PQC Edge-AI Split Inference

This is a public, demo/review version of a larger project. Datasets, trained weights, and the server-GPU energy campaign are private, so the public results here come from a smaller subset and some numbers differ from the full private work. The trusted public source is results/verified_summary.csv.

Overview

This repository contains a sanitized implementation of an AttentionGrid split-inference benchmark for edge video analytics under classical TLS and post-quantum transport modes. The benchmark compares full-frame cloud inference against AttentionGrid tile selection on an edge client, with optional application-layer classical or PQC hybrid protection around each request.

Quick Technical Review

• Inspect the AttentionGrid selection and adaptive-encoding path in src/attention_gridv2.py.
• Follow request framing and cryptographic protection through pqc_protocol.py and pqc_crypto.py.
• Check the verified public results with scripts/verify_public_results.py; no private dataset or model weights are required.

What Problem This Solves

Edge devices often have limited compute and uplink bandwidth, while remote inference adds network and cryptographic cost. This project evaluates whether co-optimizing attention-based tile selection and transport security can reduce uploaded data while preserving detection quality and verifiable transport behavior.

System Architecture

Core paths:

• experiments/run_split_inference_benchmark.py: experiment runner for edge/cloud and baseline/AttentionGrid modes.
• src/attention_gridv2.py: saliency, grid/tile selection, and adaptive encoding.
• src/network/: HTTPS inference server, client protocol, and classical/PQC transport helpers.
• src/eval_tools/: dataset discovery, detection metrics, unique-object recall, and system telemetry.
• configs/: scene-level AttentionGrid configuration and curated PQC sweep pairs.
• scripts/: Jetson smoke-test runner, public PQC sweep runner, and public result verifier.

Hardware and Software Requirements

Tested project runs targeted:

• Edge client: NVIDIA Jetson Orin NX class device.
• Ground station: CUDA-capable Linux host for YOLO inference.
• Python 3.10+.
• NVIDIA CUDA/PyTorch stack appropriate for the target Jetson or ground-station platform.
• git, cmake, and build tools for liboqs-python.

The public repository does not include datasets, trained weights, generated labels, raw benchmark runs, certificates, or private environment files.

Build Instructions

Run from the repository root.

python3 -m venv .venv
source .venv/bin/activate

# On Jetson, install the NVIDIA-supported PyTorch/torchvision wheel first.
pip install -r requirements.txt

Generate a self-signed server certificate on the ground station:

./src/network/generate_cert.sh GROUND_STATION_IP

Copy src/network/certs/server.crt to the edge client if the server and client are on separate machines.

Run and Reproduce

Expected dataset layout under the repository root:

ua_detrac/content/UA-DETRAC/DETRAC_Upload/images/val/...
ua_detrac/content/UA-DETRAC/DETRAC_Upload/labels/val/...
ua_detrac/content/UA-DETRAC/DETRAC_Upload/labels_with_ids/val/...
mot17/MOT17/train/...

Start the ground-station server:

PYTHONPATH=src python -m network.server \
  --host 0.0.0.0 \
  --port 8443 \
  --weight yolo11s.pt \
  --device cuda \
  --no-prompt

Run a Jetson Orin smoke test for AttentionGrid + PQC:

REMOTE_URL=https://GROUND_STATION_IP:8443/infer \
REMOTE_CAFILE=src/network/certs/server.crt \
MAX_SEQUENCES=1 \
./scripts/run_jetson_orin_experiment.sh

Additional reproduction notes are in docs/reproduction.md.

Run a specific public-result operating point:

PYTHONPATH=src python experiments/run_split_inference_benchmark.py \
  --dataset ua_detrac \
  --mode cloud_ag \
  --crypto-mode pqc \
  --pqc-kem ML-KEM-768 \
  --pqc-sig ML-DSA-65 \
  --remote-url https://GROUND_STATION_IP:8443/infer \
  --remote-cafile src/network/certs/server.crt \
  --non-interactive

Run the curated PQC sweep used by the public summary:

REMOTE_URL=https://GROUND_STATION_IP:8443/infer \
REMOTE_CAFILE=src/network/certs/server.crt \
DATASET=ua_detrac \
MODE=cloud_ag \
./scripts/run_public_pqc_sweep.sh

Key Verified Results

The released summary is results/verified_summary.csv. It contains 340 rows: 85 transport configurations for each dataset/mode combination across UA-DETRAC cloud baseline, UA-DETRAC cloud AttentionGrid, MOT17 cloud baseline, and MOT17 cloud AttentionGrid. Verification status, signature/decryption failure counts, payload hash checks, timing, bandwidth, accuracy, and telemetry metrics are included.

Verify the public summary:

python scripts/verify_public_results.py

Selected rows copied exactly from results/verified_summary.csv:

Dataset	Mode	Transport	Sequences	Frames	FPS	mAP_50	Unique recall	Net upload MB	Avg RTT ms	Avg PQ sign ms	Verification
ua_detrac	cloud_baseline	classical	40	56340	14.575426082404524	0.7887680386454994	0.995519413376781	341.5926197052002	38.531959471702606	0.0	True
ua_detrac	cloud_baseline	ML-KEM-768 + ML-DSA-65	40	56340	14.028069088272435	0.7887680386454994	0.995519413376781	349.4945358276367	39.89171083152946	2.821906350675323	True
ua_detrac	cloud_ag	classical	40	56340	26.60938899096889	0.7023915756083356	0.9950082440024337	83.27949299812317	31.333781755068173	0.0	True
ua_detrac	cloud_ag	ML-KEM-768 + ML-DSA-65	40	56340	27.70964249435574	0.7023915756083356	0.9950082440024337	86.36703839302064	30.21689020118606	1.6222213314265477	True
mot17	cloud_baseline	classical	7	5316	7.7146948806774605	0.7380300020467594	0.9524993578714935	375.42105538504467	108.11250126131462	0.0	True
mot17	cloud_baseline	ML-KEM-768 + ML-DSA-65	7	5316	7.62768436135504	0.7380300020467594	0.9524993578714935	379.6815414428711	108.79375112063074	2.9735494385733494	True
mot17	cloud_ag	classical	7	5316	11.37235775276736	0.6639733477363562	0.9334837951481472	142.3633279800415	72.14300010911329	0.0	True
mot17	cloud_ag	ML-KEM-768 + ML-DSA-65	7	5316	11.457471058242273	0.6639733477363562	0.9334837951481472	144.95119789668493	73.59846595946367	2.1176792730235716	True

Expected Output

Each experiment writes to:

runs/<timestamp>_<mode>_<dataset>/

Expected files include run_config.json, per_sequence_results.csv, summary.csv, and, for sweeps, master_summary.csv plus one subdirectory per transport configuration.

Limitations

• Full reproduction requires the supported datasets, YOLO weights, certificates, and suitable edge/server hardware.
• The public release includes verified summaries, not raw internal benchmark directories.
• Jetson installation depends on the NVIDIA-provided PyTorch/CUDA package set for the specific JetPack version.
• AttentionGrid behavior depends on scene configuration files in configs/; changing dataset splits or camera scenes requires re-tuning or adding scene configs.

Sanitization Note

This is a sanitized academic/portfolio version. Proprietary research artifacts, internal documents, unpublished manuscripts, and private datasets are excluded.