ehw-fit/obfax

IP Theft Detection — Paired Approximate Multiplier Dataset

Dataset Summary

This dataset supports the detection of intellectual property (IP) theft in approximate circuits. Each sample is a pair of 8-bit approximate multipliers represented as 256×256 error heat maps. The binary label indicates whether one multiplier was derived (obfuscated) from the other, enabling a Siamese-network-based similarity learning task.

Dataset Structure

Files

File	Split	Description
pairs_data.pkl.gz	all	Full paired dataset
train_data.pkl.gz	train	Training split
val_data.pkl.gz	val	Validation split
test_data.pkl.gz	test	Test split

All files are gzip-compressed pandas DataFrames (.pkl.gz).

Columns

Column	Type	Description
chrom1	string / array	CGP chromosome encoding of multiplier 1
chrom2	string / array	CGP chromosome encoding of multiplier 2
multiplier1	identifier	Identifier of multiplier 1
multiplier2	identifier	Identifier of multiplier 2
same	int (0/1)	Label. 1 if mult2 was derived from mult1 (IP theft), 0 otherwise
feat1	ndarray (256, 256)	Error heat map for multiplier 1
feat2	ndarray (256, 256)	Error heat map for multiplier 2

Error Heat Map

Each heat map is the arithmetic error surface of an 8-bit approximate multiplier over all 256×256 input combinations:

feat[a, b] = circuit(a, b) − a × b      for a, b ∈ {0, …, 255}

The circuit is evaluated using UnsignedCGPCircuit from the ariths_gen library:

from ariths_gen.core.cgp_circuit import UnsignedCGPCircuit

c = UnsignedCGPCircuit(chrom, [8, 8])
out = c(a, b)
diff = out - a * b          # element at position (a, b) in the heat map

Label Definition

A pair (mult1, mult2) is labelled same = 1 when mult2 was produced by obfuscating mult1 (e.g. gate-level modifications that preserve approximate behaviour while disguising provenance). Pairs with same = 0 are unrelated multipliers sampled from the same design space.

Loading the Dataset

import pandas as pd

df = pd.read_pickle("data/pairs_data.pkl.gz")

# Access a heat map pair
feat1 = df["feat1"].iloc[0]   # shape (256, 256)
feat2 = df["feat2"].iloc[0]   # shape (256, 256)
label = df["same"].iloc[0]    # 0 or 1

Intended Use

• Training and evaluating Siamese / contrastive learning models for circuit fingerprinting
• Benchmarking approximate-circuit similarity metrics
• Research on IP protection in approximate computing

Out-of-Scope Use

• Direct deployment as a production IP-theft detection system without further validation
• Generalisation claims beyond 8-bit unsigned multipliers without additional experiments

Source & Provenance

Circuits are Cartesian Genetic Programming (CGP) encodings of 8-bit unsigned approximate multipliers generated with the ariths_gen framework. Obfuscated variants were derived programmatically from originals to simulate IP theft scenarios.

Citation

@inproceedings{sekanina2026obfax,
  author    = {Lukas Sekanina and Vojtech Mrazek},
  title     = {{ObfAx}: Obfuscation and {IP} Piracy Detection in Approximate Circuits},
  booktitle = {Proceedings of the Great Lakes Symposium on VLSI (GLSVLSI)},
  year      = {2026},
  doi       = {10.1145/3787109.3815215}
}