Trustworthy Anomaly Detection: A Survey (2202.07787v1)

Abstract: Anomaly detection has a wide range of real-world applications, such as bank fraud detection and cyber intrusion detection. In the past decade, a variety of anomaly detection models have been developed, which lead to big progress towards accurately detecting various anomalies. Despite the successes, anomaly detection models still face many limitations. The most significant one is whether we can trust the detection results from the models. In recent years, the research community has spent a great effort to design trustworthy machine learning models, such as developing trustworthy classification models. However, the attention to anomaly detection tasks is far from sufficient. Considering that many anomaly detection tasks are life-changing tasks involving human beings, labeling someone as anomalies or fraudsters should be extremely cautious. Hence, ensuring the anomaly detection models conducted in a trustworthy fashion is an essential requirement to deploy the models to conduct automatic decisions in the real world. In this brief survey, we summarize the existing efforts and discuss open problems towards trustworthy anomaly detection from the perspectives of interpretability, fairness, robustness, and privacy-preservation.

• The paper outlines the need to go beyond accuracy by integrating interpretability, fairness, robustness, and privacy into anomaly detection systems.
• The paper surveys both shallow and deep model approaches, highlighting how current methods address individual trustworthiness dimensions.
• The paper advocates for future research on unified evaluation metrics, integrated frameworks, and benchmark datasets to enhance trustworthiness in sensitive applications.

A Survey on Trustworthy Anomaly Detection

The paper "Trustworthy Anomaly Detection: A Survey" significantly contributes to the growing body of research focused on anomaly detection within the broader context of machine learning. Authored by Shuhan Yuan and Xintao Wu, this survey highlights the necessity of trustworthiness in anomaly detection models, acknowledging their extensive application in sensitive domains such as bank fraud detection and cyber intrusion detection. The paper delineates trustworthiness through four dimensions: interpretability, fairness, robustness, and privacy-preservation, appropriately expanding the frontier of anomaly detection research.

Trustworthiness in Anomaly Detection

The core of the paper emphasizes that detecting anomalies with only high accuracy is insufficient, especially when these models are implemented in real-world scenarios affecting human lives. The detection results must be trustworthy; this entails being understandable, fair across demographic groups, resistant to adversarial inputs, and safeguarding user privacy.

• Interpretability: The detection models should provide clear explanations for their decisions. This is particularly crucial in scenarios like fraud detection, where understanding anomalous behavior can have legal and financial ramifications.
• Fairness: Detection algorithms should not exhibit bias towards any particular group. The paper highlights the risk of disproportionate labeling of minority group data points as anomalies, thereby exacerbating existing inequalities.
• Robustness: Anomaly detection models must withstand adversarial attacks aiming to deceive the system into false positives or negatives. Robust models ensure reliability and consistency in outcomes despite external manipulations.
• Privacy-preservation: Given that anomaly detection systems often operate on sensitive data, ensuring privacy through methods like differential privacy and cryptographic techniques is critical to prevent misuse or unauthorized data access.

Survey of Existing Methodologies

The paper methodically surveys the existing literature, categorizing anomaly detection methods into shallow and deep model approaches while exploring each model type's compatibility with trustworthiness metrics. A significant portion of this discussion assesses methods such as density models, one-class models, and reconstruction models in the wish to imbue them with interpretability, fairness, robustness, and privacy-preserving features.

One insightful observation is the diversity of existing methods geared towards achieving these dimensions singularly, with most research concentrating heavily on model performance rather than holistic trustworthiness. The survey includes dense tabular analyses (as illustrated in Table 1) that contrast various approaches across multiple trustworthiness dimensions, showcasing the breadth and specificity of work carried out thus far.

Implications and Future Directions

The paper's implications resonate on both theoretical and applied levels. The survey suggests that while there has been progress in isolated aspects of trustworthiness, a unified framework that concurrently addresses all dimensions is rare but highly desirable. This understanding points towards a significant gap in current research initiatives, prompting the exploration of novel avenues in anomaly detection.

For future developments, the authors advocate for:

• Calibration of Evaluation Metrics: Enhancing the accuracy and reliability of metrics like AUROC and precision-recall in low prevalence settings to better gauge model performance.
• Design of Unified Frameworks: A multidisciplinary approach to designing anomaly detection systems that integrate interpretability, fairness, robustness, and privacy-preservation from inception rather than as an afterthought.
• Establishment of Benchmark Datasets: Creating specialized datasets to evaluate the intersectionality of these dimensions can foster comprehensive methodological improvements.

In conclusion, "Trustworthy Anomaly Detection: A Survey" articulates a comprehensive vision for the evolution of anomaly detection research. It grounds the discussion in real-world applications, urging advancements that align AI's computational strengths with nuanced societal values. This serves as a critical call to the research community to prioritize trustworthiness as a standard criterion in future model development and evaluation.