Trust Filter: Runtime Hardware Trojan Detection in Behavioral MPSoCs

Abstract

This work introduces a runtime system level method to detect hardware Trojan in third party behavioral intellectual properties (3PBIPs). Most of the HW Trojan detection techniques either rely on golden Trojanfree (trusted) models, which are compared to the suspected model, or source IPs with the same functionality from different vendors. In the case of BIPs, this is extremely hard, as it is a market still in its infancy. Moreover, it is very difficult to find HW Trojan at the system level as the trigger condition might be visible only when the system is fully functional. Thus, this work proposes the inclusion of a small HW Trojan detection circuit called trust filters to detect HW Trojan at runtime. With the help of cycle-accurate simulation model, it is possible to fine tune these filters so that the overall system has no performance degradation. This can be achieved by exploiting the slack time between the time that a slave returns the data to the master and the time that the master sends new data to the slave. The advantages of using C-based design are multi-fold: (i) It allows the generation of fast cycle-accurate models to measure the exact slack of each BIP mapped as a loosely coupled Hardware Accelerator (HWAcc) slave and (ii) its ability to build the complete SoC using synthesizable Application Programming Interfaces (APIs) and hence allowing the fine tuning of these trust filters. Experimental results show that our proposed architecture is very efficient leading to no performance penalties in many cases and has very small area overhead.Keywords High-level synthesis · Hardware trojan · Behavioral MPSoC · Third party behavioral IP (3PBIP)

Hardware Trojan detection and localization based on local detectors

“…This idea can offer more than 100% area and power consumption redundancy. In this regard, other studies focused on IP level, including [10]. There are additional methods to detect Trojans based on run-time techniques.…”

“…Overhead (area/power) This work 2017 Logical detectors 0.5%/10% 0.5%/8% [10] 2017 Trust Filter -1.7%/6.45% [18] 2016 Weighted random pattern 10.26% 9.5% [16] 2016 TPAD (hardware Trojan 7.4%/165% -prevention and detection) [11] 2014 Current sensors 0.5%/8% 0.1%/2% [17] 2014 Insertion MUX 0.3%/32% 0.6%/45% [14] 2014 Delay fingerprint 19%/26% - [15] 2013 Camouflaging 1 5.5X 4X [3] 2012 Thermal sensors 10%/15% -1 5.5X implies a 5.5-fold increase of overhead, 4X implies a four-fold increase of overhead.…”

“…In this paper, the overhead had a normal value. The figures reported for different detection methods [3,10,11,[14][15][16][17][18] are shown in Table 3.…”