An intriguing property of deep neural networks (DNNs) is their inherent vulnerability to adversarial inputs, which significantly hinder the application of DNNs in security-critical domains. Despite the plethora of work on adversarial attacks and defenses, many important questions regarding the inference behaviors of adversarial inputs remain mysterious. This work represents a solid step towards answering those questions by investigating the information flows of normal and adversarial inputs within various DNN models and conducting in-depth comparative analysis of their discriminative patterns. Our work points to several promising directions for designing more effective defense mechanisms.

Many of today's machine learning (ML) systems are built by reusing an array of, often pre-trained, primitive models, each fulfilling distinct functionality (e.g., feature extraction). The increasing use of primitive models significantly simplifies and expedites the development cycles of ML systems. Yet, because most of such models are contributed and maintained by untrusted sources, their lack of standardization or regulation entails profound security implications, about which little is known thus far. In this paper, we demonstrate that malicious primitive models pose immense threats to the security of ML systems. We present a broad class of model-reuse attacks wherein maliciously crafted models trigger host ML systems to misbehave on targeted inputs in a highly predictable manner. By empirically studying four deep learning systems (including both individual and ensemble systems) used in skin cancer screening, speech recognition, face verification, and autonomous steering, we show that such attacks are (i) effective - the host systems misbehave on the targeted inputs as desired by the adversary with high probability, (ii) evasive - the malicious models function indistinguishably from their benign counterparts on non-targeted inputs, (iii) elastic - the malicious models remain effective regardless of various system design choices and tuning strategies, and (iv) easy - the adversary needs little prior knowledge about the data used for system tuning or inference. We provide analytical justification for the effectiveness of model-reuse attacks, which points to the unprecedented complexity of today's primitive models. This issue thus seems fundamental to many ML systems. We further discuss potential countermeasures and their challenges, which lead to several promising research directions.

As one critical task in the data analysis pipeline, data cleaning is notoriously human labor-intensive and error-prone. Knowledge base-assisted data cleaning has proved a powerful tool for finding and fixing data defects; however, its applicability is inevitably bounded by the natural limitations of knowledge bases. Meanwhile, although a vast number of knowledge sources exist in the form of free-text corpora (e.g., Wikipedia), transforming them into formats usable by existing data cleaning tools can be prohibitively costly and error-prone, if not at all impossible. Here, we present DeepClean, the first end-to-end data cleaning framework powered by free-text knowledge sources. At a high level, DeepClean leverages a knowledge source through its question-answering (QA) interface and achieves high-quality cleaning via iterative question asking. Specifically, DeepClean detects and repairs data defects in three stages: (i) Pattern extraction - it automatically discovers the semantic types of the data attributes as well as their correlations; (ii) Question generation - it translates each data tuple into a minimal set of validation questions; (iii) Completion and repair - by checking the answers returned by the knowledge source against the data values, it identifies erroneous cases and suggests possible fixes. Through extensive empirical studies, we demonstrate that DeepClean is applicable to a range of domains, and can effectively repair a variety of data defects, highlighting data cleaning powered by free-text knowledge sources as a promising direction for future research.