Tutorial 63: Hardware Fault Resilience Testing

Test how your SNN degrades under hardware faults before deployment. The resilience suite injects stuck-at faults, bit flips, dead synapses, and stochastic computing biases, measures accuracy degradation per fault rate, and identifies the most vulnerable layer.

Why Resilience Testing

FPGA and ASIC hardware develops faults over time: radiation-induced bit flips (space, medical), wear-out (automotive), manufacturing defects. An SNN that drops from 97% to 50% accuracy with 1% stuck weights is not deployable. One that drops to 95% is.

Quick Start

import numpy as np
from sc_neurocore.resilience import FaultResilienceSuite
from sc_neurocore.resilience.fault_suite import FaultType

rng = np.random.default_rng(42)

# Your evaluation function (returns accuracy)
def my_eval(weights):
    # Replace with actual inference + accuracy measurement
    # Returns float in [0, 1]
    return 0.95 - np.mean([np.abs(w).mean() for w in weights]) * 0.1

model_weights = [
    rng.standard_normal((128, 784)).astype(np.float32) * 0.05,
    rng.standard_normal((10, 128)).astype(np.float32) * 0.1,
]

suite = FaultResilienceSuite(eval_fn=my_eval, weights=model_weights)

# Sweep stuck-at-zero faults at increasing rates
report = suite.sweep(FaultType.STUCK_AT_ZERO, rates=[0.01, 0.05, 0.1, 0.2])
print(report.summary())
# Fault: STUCK_AT_ZERO
# Rate 0.01: accuracy 94.8% (Δ -0.2%)
# Rate 0.05: accuracy 93.1% (Δ -1.9%)
# Rate 0.10: accuracy 89.7% (Δ -5.3%)
# Rate 0.20: accuracy 81.2% (Δ -13.8%)
# Critical threshold: ~8% fault rate for >5% accuracy drop

Full Audit

Test all fault types across all layers:

full = suite.full_audit()
print(f"Most vulnerable layer: {full.most_vulnerable_layer()}")
print(f"Most damaging fault: {full.most_damaging_fault()}")
print(f"Overall resilience score: {full.resilience_score()}/100")

# Per-layer, per-fault breakdown
for layer, faults in full.layer_results.items():
    for fault_type, result in faults.items():
        print(f"  {layer} × {fault_type}: Δ={result.accuracy_drop:.1%} at 5% rate")

Fault Types

Type	Effect	SC-Specific?	Severity
STUCK_AT_ZERO	Weights clamped to 0	No	High (silences synapses)
STUCK_AT_ONE	Weights clamped to 1	No	High (saturates activity)
WEIGHT_BIT_FLIP	Random bit flipped in Q8.8	No	Medium (depends on bit position)
DEAD_SYNAPSE	Entire connections zeroed	No	High (structural damage)
NOISY_MEMBRANE	Gaussian noise on membrane	No	Low (SNNs are noise-tolerant)
BITSTREAM_BIAS	SC probability bias toward 0.5	Yes	Medium (degrades SC precision)

The BITSTREAM_BIAS fault is unique to stochastic computing — it models LFSR correlation that causes bitstream probabilities to drift toward 0.5. No other framework tests for this.

Hardening Strategies

If resilience testing reveals vulnerability, SC-NeuroCore provides mitigation:

Strategy	How	Accuracy Cost
Mismatch-aware training (Tutorial 48)	Inject faults during training	<1%
Weight redundancy	Duplicate critical synapses	0% (more resources)
Error-correcting codes	SECDED on weight BRAMs	0% (adds parity bits)
Threshold homeostasis (Tutorial 68)	Auto-regulate after faults	~1%

Comparison

Feature	SC-NeuroCore	Lava	snnTorch
Fault injection	6 types	No	No
Per-layer vulnerability	Yes	No	No
SC-specific faults	Yes	No	No
Automated sweep	Yes	No	No
Resilience scoring	Yes	No	No

References

• Schuman et al. (2022). "Resilience and Robustness of Spiking Neural Networks for Neuromorphic Systems." IJCNN 2022.
• El-Sayed et al. (2018). "Spiking neural network robustness to permanent hardware faults." Neural Computing and Applications.