Side-Channel Encoding

SC-NeuroCore includes analytic tooling for stochastic-computing bitstreams whose switching activity should be inspected before deployment. The current workflow generates deterministic activity-balanced streams, optional dummy streams, analytic leakage-proxy reports, deploy/evidence manifests, and ROM-style HDL hooks for precomputed bitstreams.

The evidence boundary is analytic_simulation_only.

What The Evidence Means

The side-channel workflow can report:

• per-stream transition counts and transition rates;
• class-conditioned activity proxy values;
• baseline-versus-protected class-mean gap reduction;
• dummy-stream overhead ratio;
• protected bitstream count;
• total dummy streams inserted;
• generated HDL hook metadata.

The side-channel workflow does not provide:

• no physical power measurement;
• no physical thermal measurement;
• no DPA-resistance claim;
• no silicon-security claim.

Treat these artefacts as design-time evidence for stochastic-encoding review. Board measurements, oscilloscope or power-rail traces, thermal captures, place-and-route timing, and adversarial lab validation are separate evidence classes and must be recorded separately.

Python Workflow

Use the activity-balanced encoder for deterministic, probability-preserving bitstream generation:

Python

from sc_neurocore.security import (
    ThermalSCEncodingConfig,
    encode_activity_balanced_probability,
)

record = encode_activity_balanced_probability(
    0.25,
    ThermalSCEncodingConfig(
        bitstream_length=64,
        seed=5,
        dummy_streams_per_record=1,
        max_dummy_overhead_ratio=1.0,
    ),
)

The returned record contains the payload bitstream, dummy bitstreams, realised probability, switching-activity summary, dummy-stream count, seed domain, and the analytic_simulation_only evidence boundary.

Benchmark Report

Generate a canonical JSON benchmark report from the command line:

Bash

PYTHONPATH=src python tools/side_channel_benchmark.py \
  --output side_channel_benchmark.json \
  --probabilities 0.25,0.5 \
  --labels 0,1 \
  --bitstream-length 64 \
  --seed 5 \
  --dummy-streams-per-record 1 \
  --max-dummy-overhead-ratio 1.0

The report contains:

• schema_version;
• evidence_boundary;
• deploy_manifest;
• baseline class-activity proxy;
• protected class-activity proxy;
• max class-mean gap reduction;
• boundary notes;
• per-sample records.

The deploy_manifest section records the evidence class, benchmark artifact path, security parameters, overhead measurements, and the same physical evidence boundary notes.

HDL Hook

Generate a ROM-style Verilog hook and hook manifest from the same protected encoding configuration:

Bash

PYTHONPATH=src python tools/side_channel_hdl_emit.py \
  --verilog-output rtl/sc_side_channel_hook.v \
  --manifest-output rtl/sc_side_channel_hook_manifest.json \
  --module-name sc_side_channel_hook \
  --probability 0.25 \
  --bitstream-length 64 \
  --seed 5 \
  --dummy-streams-per-record 1 \
  --max-dummy-overhead-ratio 1.0

The emitted HDL exposes payload and dummy stream bits by sample_index. It is a transport hook for precomputed analytic bitstreams. It is not a replacement for hardware leakage measurement or silicon validation.

Review Checklist

Before using a side-channel report in a deployment packet, check that:

• the report has evidence_boundary: analytic_simulation_only;
• deploy_manifest.security_parameters records bitstream length, seed, rotation stride, dummy stream count, and dummy overhead budget;
• deploy_manifest.overhead_measurements records the dummy overhead ratio and inserted dummy stream count;
• generated HDL hook manifests point to the Verilog file actually emitted;
• physical measurement claims are absent unless separate raw measurement artefacts exist.