PAPER: N/A
================================================================================

REVIEW
--------------------------------------------------------------------------------
Okay, I need to write a detailed peer review for this paper introducing APBench. Let me start by summarizing the paper's claims and methods. The paper presents APBench as a benchmark for evaluating availability poisoning attacks and defenses. It includes various attacks, defenses, and datasets, and conducts experiments across different models and poisoning rates. The main contributions are the benchmark itself, comprehensive evaluations, and insights into the effectiveness of different methods.

Next, the strengths. The paper has several strengths. First, it's comprehensive, covering both supervised and unsupervised attacks along with multiple defenses. Second, the evaluation is thorough, testing across multiple datasets and model architectures. Third, the inclusion of qualitative analysis tools like t-SNE and Grad-CAM adds depth. Fourth, the open-source nature and detailed documentation in appendices are commendable.

Now, the weaknesses. The first concern is the lack of comparison to existing benchmarks. The paper mentions RobustBench and Backdoorbench but doesn't clearly explain how APBench differs. Another issue is the absence of statistical validation for results—no confidence intervals or p-values. Additionally, the explanation of why $\ell_0$-bounded attacks are more robust against defenses is unclear. Also, the choice of baselines like CutOut and MixUp as defenses seems arbitrary without justification. Lastly, the paper doesn't address the scalability or adaptability of the benchmark to other data modalities beyond images.

For questions to the authors, I need to ask specific points. For example, how APBench differs from existing benchmarks, the statistical significance of results, the rationale behind $\ell_0$ robustness, the selection of baseline defenses, and plans for expanding to other modalities.

Limitations not addressed include the lack of exploration into semi-supervised scenarios, insufficient discussion on the trade-offs between attack stealth and effectiveness, and minimal analysis of the impact of different training schedules on defense performance.

In terms of soundness, I'd rate it a 3 (good) because the methodology is solid but lacks some statistical rigor. Contribution is around 3 as well since it's a timely benchmark but could have deeper analysis. Confidence is 4 because the work is credible but has gaps. The rating would be 8 (accept) due to the overall usefulness despite the shortcomings.

Now, putting it all together in the required structure.
</think>

### Summary Of The Paper  
The paper introduces **APBench**, a benchmark for evaluating **availability poisoning (APA)** attacks and defenses in deep learning. It aggregates 9 supervised APA attacks (e.g., EM, REM, LSP), 2 unsupervised APA attacks (UCL, TUE), 9 defenses (e.g., JPEG compression, adversarial training), and 4 data augmentation techniques. Experiments are conducted across 4 datasets (CIFAR-10, CIFAR-100, SVHN, ImageNet-subset) and 5 model architectures (ResNet-18, ResNet-50, MobileNetV2, DenseNet-121, ViT). Key findings highlight the ineffectiveness of $\ell_p$-bounded attacks against advanced defenses like UEraser-Max and AVATAR, the robustness of $\ell_0$-based attacks (e.g., OPS) against preprocessing, and the limited success of defenses against transferable attacks. The paper concludes that APBench serves as a tool to accelerate research on APA and privacy-preserving defenses.

---

### Strengths  
1. **Comprehensive Coverage**: APBench includes a wide array of APA attacks (both supervised and unsupervised) and defenses, enabling systematic comparisons (Sections 2–4, Tables 1–2).  
2. **Thorough Evaluation**: The benchmark evaluates attack-defense interactions across diverse datasets, poisoning rates (20%–80%), and model architectures (Tables 3–6, Figures 3–4).  
3. **Qualitative Insights**: Tools like t-SNE, Grad-CAM, and Shapley value maps are employed to analyze the impact of poisoning (Section 4, Figures 6–8).  
4. **Open-Source Accessibility**: Code and detailed hyperparameters are provided (Appendix C, Table 11), enhancing reproducibility.  

---

### Weaknesses  
#### 1. **Insufficient Comparison to Existing Benchmarks** *(Severe)*  
**Problem**: The paper claims APBench is the "first" benchmark for APA (Section 1), but it does not clarify how it differs from existing benchmarks like RobustBench [4] (for adversarial robustness) or Backdoorbench [42] (for backdoor attacks).  
**Location**: Section 1 (“Related Benchmarks”) and Table 16 (licenses/code sources).  
**Recommendation**: Explicitly compare APBench’s scope, methodologies, and evaluation metrics to prior benchmarks. Highlight unique contributions (e.g., unsupervised APA support, poisoning rate variability).  

#### 2. **Lack of Statistical Validation** *(Major)*  
**Problem**: Results (e.g., Table 3, where EM+JPEG achieves 90.29% accuracy on ResNet-18) are presented without confidence intervals or p-values to assess statistical significance.  
**Location**: Tables 3–6, Section 4.  
**Recommendation**: Include error bars or repeated experiments (e.g., 3–5 runs) to validate the robustness of observed trends.  

#### 3. **Ambiguous Explanation of $\ell_0$-Bound Attack Robustness** *(Moderate)*  
**Problem**: The paper claims $\ell_0$-bounded attacks (e.g., OPS) are "robust against a plethora of defenses" (Section 4), but no formal justification is provided for why $\ell_0$ perturbations resist preprocessing or adversarial training.  
**Location**: Section 4, paragraph discussing $\ell_0$-bound attacks.  
**Recommendation**: Add a subsection explaining the mathematical or empirical basis for this robustness (e.g., resistance to JPEG compression, adversarial training failure modes).  

#### 4. **Arbitrary Choice of Baseline Defenses** *(Minor)*  
**Problem**: CutOut, MixUp, and CutMix are listed as baselines (Table 2), but these are standard data augmentation techniques, not specifically designed for APA. Their inclusion as "defenses" lacks motivation.  
**Location**: Table 2, Section 4.  
**Recommendation**: Clarify whether these are included for completeness or as active defenses. Compare their performance to domain-specific defenses (e.g., AVATAR) to contextualize their relevance.  

---

### Questions For The Authors  
1. **How does APBench differ from existing benchmarks like RobustBench [4] or Backdoorbench [42] in terms of scope, methodology, or evaluation criteria?**  
2. **Were the reported test accuracies (e.g., Table 3) computed over multiple runs (≥3)? What is the variance across runs?**  
3. **What is the theoretical or empirical basis for the claim that $\ell_0$-bounded attacks (e.g., OPS) are robust against preprocessing and adversarial training?**  
4. **Why are CutOut, MixUp, and CutMix included as "defenses" in Table 2? Are they intended as baselines for comparison to domain-specific defenses like AVATAR?**  
5. **Does APBench support extension to non-image modalities (e.g., text, audio)? If not, what are the barriers to such generalization?**  

---

### Limitations Not Addressed By The Authors  
1. **Limited Exploration of Semi-Supervised Scenarios**: The paper focuses on fully supervised or unsupervised APA but does not examine hybrid scenarios (e.g., semi-supervised learning with poisoned data).  
2. **Neglect of Trade-Offs Between Stealth and Effectiveness**: While Fig. 1 highlights stealthiness, the paper does not quantify how perturbation visibility correlates with attack success rates.  
3. **Minimal Discussion of Training Schedule Impact**: Defense performance varies with training epochs (e.g., 200 vs. 300 in Table 14), but the paper does not analyze how training duration affects robustness.  

---

### Soundness: **3 (Good)**  
The methodology is well-defined, but the lack of statistical validation and unclear theoretical foundations for $\ell_0$-attack robustness reduce confidence in the conclusions.  

### Contribution: **3 (Good)**  
APBench addresses a timely need for APA benchmarking but could strengthen its novelty by contrasting with existing work and addressing scalability limitations.  

### Confidence: **4 (High)**  
The paper is technically sound and provides a valuable resource, though gaps in statistical rigor and theoretical justification warrant clarification.  

### Rating: **8 (Accept)**  
APBench makes a meaningful contribution to the field by offering a standardized evaluation framework for APA. Minor improvements in clarity and statistical reporting would elevate its impact.  

### Brief Justification For Rating  
The paper introduces a comprehensive benchmark for APA, filling a critical gap in the literature. While it has room for improvement in statistical validation and theoretical grounding, its open-source implementation, broad coverage of attacks/defenses, and qualitative analysis tools justify acceptance.

================================================================================
