2D-OOB: Attributing Data Contribution Through Joint Valuation Framework

TL;DR

2D-OOB is proposed, an out-of-bag estimation framework for jointly determining helpful (or detrimental) samples as well as the particular cells that drive them that show promising results in detecting and rectifying fine-grained outliers at the cell level, and localizing backdoor triggers in data poisoning attacks.

Abstract

Data valuation has emerged as a powerful framework for quantifying each datum's contribution to the training of a machine learning model. However, it is crucial to recognize that the quality of cells within a single data point can vary greatly in practice. For example, even in the case of an abnormal data point, not all cells are necessarily noisy. The single scalar score assigned by existing data valuation methods blurs the distinction between noisy and clean cells of a data point, making it challenging to interpret the data values. In this paper, we propose 2D-OOB, an out-of-bag estimation framework for jointly determining helpful (or detrimental) samples as well as the particular cells that drive them. Our comprehensive experiments demonstrate that 2D-OOB achieves state-of-the-art performance across multiple use cases while being exponentially faster. Specifically, 2D-OOB shows promising results in detecting and rectifying fine-grained outliers at the cell level, and localizing backdoor triggers in data poisoning attacks.

Figures (9)

• Figure 1: Comparison of data valuation and joint valuation. (a) Data valuation evaluates the quality of individual data points, whereas (b) joint valuation evaluates the quality of individual cells. Both panels illustrate the same hypothetical dataset, while darker colors indicate higher quality or importance. As illustrated in panel (a), data valuation can only identify that the third and fifth data points are of low quality, but it lacks further feature-level attribution. This limitation may result in discarding the entire data point, even when only certain cells are problematic. In contrast, joint valuation provides a finer level of attribution than data valuation and aims to reveal how individual features contribute to data values. As shown in panel (b), the joint valuation framework can identify outlier cells (highlighted by blue boxes), such as $-1$ in "Income" and $100$ in "Education", providing detailed interpretations of data values.
• Figure 2: Cell-level outlier detection rate curves for 2D-OOB, 2D-KNN, and Random. The x-axis represents the percentage of inspected cells. The y-axis represents the detection rate, defined as the ratio of the number of detected outlier cells to the total number of outlier cells present in a dataset. The error bars show a $95\%$ confidence interval based on $30$ independent experiments. We examine the cells in ascending order, starting from those with the lowest values, and thus a curve closer to the left-top corner indicates better performance. 2D-OOB efficiently detects the majority of outlier cells by examining only a small fraction of the total cells, while 2D-KNN and Random require scanning nearly all the cells.
• Figure 3: Cell fixation experiment results (test accuracy curves) for 2D-OOB, 2D-KNN, and Random. We replace cells with their ground-truth annotations, starting with those cells assigned the lowest valuations. The results for $6$ datasets are presented, and additional results for other datasets are available in Appendix \ref{['sec:additional fixation']}. We conduct $30$ independent trials and report the average results. A higher curve indicates better performance. 2D-OOB demonstrates a superior capability in accurately identifying and rectifying cell-level outliers.
• Figure 4: Backdoor trigger detection rate curves for 2D-OOB, 2D-KNN, and Random. Panels A (top) and B (bottom) correspond to the Trojan square and BadNets square, respectively. We inspect the cells within each poisoned sample in descending order of their valuation scores. The detection rate curve shows the average detection rate across all poisoned samples, with error bars representing a $95\%$ confidence interval based on $15$ independent runs. 2D-OOB demonstrates superior performance in detecting the cells implanted with triggers.
• Figure 5: Qualitative examples for 2D-OOB in the backdoor trigger detection task. Each pair of images consists of a poisoned image and its corresponding cell valuation heatmap. The color of the heatmap indicates importance, with red cells representing higher importance and blue cells representing lower importance. In the first two pairs, the class 'bird" is relabeled as "cat", while in the latter two pairs, the class "deer" is relabeled as "cat". The heatmaps clearly show that higher cell valuations are predominantly concentrated in the regions containing triggers, while areas featuring actual objects receive lower valuations. This pattern suggests that 2D-OOB effectively captures the triggers as the impactful features responsible for the misclassification of the poisoned samples.

Theorems & Definitions (4)

• Proposition 3.1
• proof
• Proposition D.1
• proof