Opportunistic Target Selection

Early Directional Commitment for Query-Efficient Black-Box Adversarial Attacks

Overview

Opportunistic Target Selection (OTS) is a lightweight wrapper for score-based black-box adversarial attacks that lack implicit target tracking. It runs the attack in untargeted mode for a short exploration phase, then switches to a targeted objective against whichever non-true class currently leads. OTS acts as a margin-loss surrogate: it reduces class drift for probability- or cross-entropy-based random-search attacks without requiring architectural modification, gradient access, or a priori target-class knowledge.

Across three score-based attacks and five standard ImageNet classifiers (4,500 runs), OTS closely tracks oracle targeting on drift-prone attacks, with gains up to +27 pp in success rate and 43% relative reduction in censored-mean iterations on ResNet-50. It is redundant for attacks that already provide directionality, such as Bandits or margin-loss Square Attack, and shows no benefit on adversarially-trained models where the difficulty distribution is bimodal.

See paper/main.tex for the full paper.

---

Project Structure

src/                    Core library
  attacks/              Attack implementations (SimBA, SquareAttack, Bandits)
  models/               Model loaders (torchvision, RobustBench)
  utils/                Image preprocessing & visualization
  demo/                 Gradio demonstrator app
demo/                   Demo entry point
benchmarks/             Benchmark scripts (generate CSV results)
analysis/               Analysis scripts (generate figures from CSVs)
slurm/                  HPC job scripts (CRIANN Arctic)
paper/                  LaTeX paper
results/                Benchmark CSVs and figures
data/                   ImageNet class index and demo images

---

Quick Start

1. Install dependencies

   pip install -r requirements-gpu.txt   # With GPU (NVIDIA CUDA)
   pip install -r requirements-cpu.txt   # CPU only

2. Launch the demonstrator

   python demo/launch.py

3. Access the interface

   Open http://127.0.0.1:7860 in your browser.

---

Benchmarks

Script	Description
benchmarks/benchmark.py	Multi-model benchmark: 5 standard models + 2 robust models, 3 attacks, 3 modes
benchmarks/winrate.py	ResNet-50 CDF benchmark: 100 images, 15K budget, bootstrapped CI
benchmarks/ablation_s.py	Stability threshold sweep S={2..15} on standard ResNet-50
benchmarks/ablation_s_robust.py	Stability threshold sweep on robust ResNet-50
benchmarks/ablation_naive.py	Naive fixed-iteration switching vs OTS (standard and robust)
benchmarks/margin.py	Margin vs CE loss ablation on SquareAttack
benchmarks/landscape.py	Per-iteration confidence history collection
benchmarks/theta.py	Perturbation alignment with oracle direction

python benchmarks/benchmark.py
python benchmarks/winrate.py

Analysis

Regenerate figures from benchmark CSVs:

python analysis/analyze_benchmark.py
python analysis/analyze_winrate.py
python analysis/analyze_ablation_s.py
python analysis/analyze_ablation_naive.py
python analysis/analyze_margin.py
python analysis/analyze_lockmatch.py
python analysis/analyze_oracle_beat.py