Current state-of-the-art methods for object detection rely on annotated bounding boxes of large data sets for training. However, obtaining such annotations is expensive and can require up to hundreds of hours of manual labor. This poses a challenge, especially since such annotations can only be provided by experts, as they require knowledge about the scientific domain. To tackle this challenge, we propose a domain-specific weakly supervised object detection algorithm that only relies on image-level annotations, which are significantly easier to acquire. Our method distills the knowledge of a pre-trained model, on the task of predicting the presence or absence of a virus in an image, to obtain a set of pseudo-labels that can be used to later train a state-of-the-art object detection model. To do so, we use an optimization approach with a shrinking receptive field to extract virus particles directly without specific network architectures. Through a set of extensive studies, we show how the proposed pseudo-labels are easier to obtain, and, more importantly, are able to outperform other existing weak labeling methods, and even ground truth labels, in cases where the time to obtain the annotation is limited.

Scanning Transmission Electron Microscopes (STEMs) acquire 2D images of a 3D sample on the scale of individual cell components. Unfortunately, these 2D images can be too noisy to be fused into a useful 3D structure and facilitating good denoisers is challenging due to the lack of clean-noisy pairs. Additionally, representing detailed 3D structure can be difficult even for clean data when using regular 3D grids. Addressing these two limitations, we suggest a differentiable image formation model for STEM, allowing to learn a joint model of 2D sensor noise in STEM together with an implicit 3D model. We show, that the combination of these models are able to successfully disentangle 3D signal and noise without supervision and outperform at the same time several baselines on synthetic and real data.

Efficiently integrity verification of received data requires Message Authentication Code (MAC) tags. However, while security calls for rather long tags, in many scenarios this contradicts other requirements. Examples are strict delay requirements (e.g., robot or drone control) or resource-scarce settings (e.g., LoRaWAN networks with limited battery capacity).

Through Information Reconciliation, two legitimate parties of Channel Reciprocity-based Key Generation assure that they extract the same key from local channel measurements. Current protocols exchange messages: Interactivity both causes delays and energy expenditure, and leaks information about the keying material to adversaries.We suggest non-interactive reconciliation, using a Siamese Network of CNNs that extracts reciprocal and suppresses nonreciprocal components in the measurements. Training and evaluating on real-world and synthetic data, we demonstrate that it blindly achieves higher correlation of the outputs at legitimate parties than the interactive state of the art, thus eliminating cost and information leakage at superior performance.

In Physical Layer Security, knowing the reciprocal state information of the legitimate terminals' wireless channel is considered a shared secret. Although questioned in recent works, the basic assumption is that an eavesdropper, residing more than half of a wavelength away from the legitimate terminals, is unable to even obtain estimates that are correlated to the state information of the legitimate channel. In this work, we present a Machine Learning based attack that does not require knowledge about the environment or terminal positions, but is solely based on the eavesdropper's measurements. It still successfully infers the legitimate channel state information as represented in impulse responses. We show the effectiveness of our attack by evaluating it on two sets of real world ultra wideband channel impulse responses, for which our attack predictions can achieve higher correlations than even the measurements at the legitimate channel.

In recent years, Network-on-Chip (NoC) has gained increasing popularity as a promising solution for the challenging interconnection problem in multi-processor systems-on-chip (MPSoCs). However, the interest of adversaries to compromise such systems grew accordingly, mandating the integration of security measures into NoC designs.

Local, blind synchronization of Channel Impulse Responses for key generation algorithms without leaking details to adversaries is challenging, due to noise from transmission and measurements. Extracting the channel characteristics as a common source of randomness using CIR promises notable improvements in key bit rates for Physical Layer Security, however, reciprocity requirements necessitate successful synchronization at both ends. We propose a one-dimensional Gaussian filter towards this end, which is simple, more robust, and outperforms competing approaches in almost all settings. To evaluate the quality of our approach, we assess the synchronization accuracy on extensive datasets. Experiments using data generated synthetically with the standardized IEEE 802.15 UWB channel model indicate a performance improvement of up to 31%. Performance on par or better than all alternatives in real measurements underlines its superior robustness against synchronization errors.

The advent of hardware Trojans is posing an increasing threat on complex integrated circuits. Network-on-Chip, the established communication infrastructure for many core systems-on-chip, are growing in complexity. Integration of third-party components, which are increasingly becoming valuable targets, exposes the surface for attacks through the injection of hardware Trojans. In this paper, we address active attacks on NoCs, and focus on the integrity of transmitted data. Basically, we use network coding for the transmission of data in order to increase efficiency and robustness.

Quantization, and the fact that channel characteristics are independent and identically distributed so far have received only little attention in reports about actual implementations of physical layer key generation schemes. They are merely assumed for channel reciprocity based key generation, although the secret key generation significantly relies on them. We set out to design a quantization preprocessing as well as an online quantization scheme which favours i.i.d. and uniform distribution of the generated values to achieve high entropy and key rates, and calculate the resulting mutual information between communication partners in a large, realistic measurement study. Our experiments indicate a remarkable increase in mutual information, and underline the applicability to various quantization and key generation schemes.