Search

Heterogeneous SoCs integrate FPGAs and microprocessor cores on the same fabric to accelerate applications, such as cryptography and deep learning. Since FPGAs share resources with the microprocessor cores, they can launch noncacheable synchronous DRAM (SDRAM) transactions through direct FPGA-to-microprocessor SDRAM interface. Therefore, if the FPGA 3rd party IPs (3PIPs) are malicious, they can launch rowhammer attacks on the SDRAM. Today’s countermeasures based on performance counters cannot detect these attacks because memory transactions from FPGAs do not pass through the cache. In addition, today’s countermeasures that count the frequency of activation of memory rows cannot identify the intellectual property (IP) that launches the attack from the FPGA. We present a security solution that monitors the SDRAM transactions from IPs on the FPGA to each bank of the microprocessor SDRAM through the FPGA-to-microprocessor SDRAM interface. The proposed monitor is implemented on the FPGA fabric. It can detect attempts to launch a rowhammer attack before it causes bit flips in the SDRAM. It utilizes 6.3% of the adaptive logic modules (ALMs) available in an Intel Cyclone V FPGA, when multiple IPs are monitored. [ABSTRACT FROM AUTHOR]

Imaging of a subwavelength microstructure made of a periodic grid-like finite set of circular rods is carried out from transient scattered field data in different configurations of sources and receivers. The goal is to identify the position of possibly missing rods. Time reversal is confirmed as a cheap yet efficient first-order diagnostic method even in the demanding context of a subwavelength microstructure. Tools of deep learning, expected to be valid in general circumstances if data acquired in sufficient numbers, are in parallel developed to image the microstructure. To that effect, recurrent neural networks (RNN) and convolutional neural networks (CNN) are both used. Pros and cons of all approaches are illustrated by comprehensive simulations from synthetic data computed via a finite difference time domain method (FD-TD) software carefully tailored to the microstructure model and used also to make the networks learn the microstructures. The analysis is completed from examples of laboratory-controlled data acquired on a microstructure prototype set within a microwave anechoic chamber. These examples confirm the good promises of neural networks even with rather scarce data as exemplified in a forward scattering case—fixed source and receiver antennas face each other and the microstructure is rotated between them. [ABSTRACT FROM AUTHOR]

Deep reinforcement learning (DRL) has been successfully applied to end-to-end autonomous driving, especially in simulation environments. However, common DRL approaches used in complex autonomous driving scenarios sometimes are unstable or difficult to converge. This paper proposes two approaches to improve the stability of the policy model training with as few manual data as possible. For the first approach, reinforcement learning is combined with imitation learning to train a feature network with a small amount of manual data for parameters initialization. For the second approach, an auxiliary network is added to the reinforcement learning framework, which can leverage the real-time measurement information to deepen the understanding of environment, without any guide of demonstrators. To verify the effectiveness of these two approaches, simulations in image information-based and lidar information-based end-to-end autonomous driving systems are conducted, respectively. These approaches are not only tested in the virtual game world, but also applied in Gazebo, in which we build a 3D world based on the real vehicle model of Ranger XP900 platform, the real 3D obstacle model, and the real motion constraints with inertial characteristics, so as to ensure that the trained end-to-end autonomous driving model is more suitable for the real world. Experimental results show that the performance is increased by over 45% in the virtual game world, and can converge quickly and stably in Gazebo in which previous methods can hardly converge. [ABSTRACT FROM AUTHOR]