Biblio

Training deep neural networks is a computationally expensive task. Furthermore, models are often derived from proprietary datasets that have been carefully prepared and labelled. Hence, creators of deep learning models want to protect their models against intellectual property theft. However, this is not always possible, since the model may, e.g., be embedded in a mobile app for fast response times. As a countermeasure watermarks for deep neural networks have been developed that embed secret information into the model. This information can later be retrieved by the creator to prove ownership. Uchida et al. proposed the first such watermarking method. The advantage of their scheme is that it does not compromise the accuracy of the model prediction. However, in this paper we show that their technique modifies the statistical distribution of the model. Using this modification we can not only detect the presence of a watermark, but even derive its embedding length and use this information to remove the watermark by overwriting it. We show analytically that our detection algorithm follows consequentially from their embedding algorithm and propose a possible countermeasure. Our findings shall help to refine the definition of undetectability of watermarks for deep neural networks.

Cyber Physical Systems (CPS) operating in modern critical infrastructures (CIs) are increasingly being targeted by highly sophisticated cyber attacks. Threat actors have quickly learned of the value and potential impact of targeting CPS, and numerous tailored multi-stage cyber-physical attack campaigns, such as Advanced Persistent Threats (APTs), have been perpetrated in the last years. They aim at stealthily compromising systems' operations and cause severe impact on daily business operations such as shutdowns, equipment damage, reputation damage, financial loss, intellectual property theft, and health and safety risks. Protecting CIs against such threats has become as crucial as complicated. Novel distributed detection and reaction methodologies are necessary to effectively uncover these attacks, and timely mitigate their effects. Correlating large amounts of data, collected from a multitude of relevant sources, is fundamental for Security Operation Centers (SOCs) to establish cyber situational awareness, and allow to promptly adopt suitable countermeasures in case of attacks. In our previous work we introduced three methods for security information correlation. In this paper we define metrics and benchmarks to evaluate these correlation methods, we assess their accuracy, and we compare their performance. We finally demonstrate how the presented techniques, implemented within our cyber threat intelligence analysis engine called CAESAIR, can be applied to support incident handling tasks performed by SOCs.

Organizations are experiencing an ever-growing concern of how to identify and defend against insider threats. Those who have authorized access to sensitive organizational data are placed in a position of power that could well be abused and could cause significant damage to an organization. This could range from financial theft and intellectual property theft to the destruction of property and business reputation. Traditional intrusion detection systems are neither designed nor capable of identifying those who act maliciously within an organization. In this paper, we describe an automated system that is capable of detecting insider threats within an organization. We define a tree-structure profiling approach that incorporates the details of activities conducted by each user and each job role and then use this to obtain a consistent representation of features that provide a rich description of the user's behavior. Deviation can be assessed based on the amount of variance that each user exhibits across multiple attributes, compared against their peers. We have performed experimentation using ten synthetic data-driven scenarios and found that the system can identify anomalous behavior that may be indicative of a potential threat. We also show how our detection system can be combined with visual analytics tools to support further investigation by an analyst.