5.2.3. Evaluation Results

All the of presented trojans and the sensor model were simulated in Cadence. The MTJ model was written in Verilog-A utilizing the current to H-field and H-field for resistance equations earlier mentioned in this study. Based on previous literature search, we also included a low-pass filter on the current-to-H-field equations at 100 MHz to approximately model the actual frequency response of the sensor. Simulated with the Class-E PA with trojans, tests were performed with a Wheatstone bridge configuration as suggested in [53], and the output went to an amplifier to allow for the determination of optimal gain for this sensing configuration.

The simulation results then were used as the input signals for the proposed BNN to classify the results. The dataset for evaluating the BNN classifier was generated by simulating the PA with various trojans in Cadence. The trojans themselves were tested by utilizing non-ideal switches that would be cycled on and off in the simulation. Each simulation was run at 1.5 V with process variations in fast-fast (FF), slow-slow (SS), and typical-typical (TT) process variations. Furthermore, the data were generated for temperatures of −40 ◦C, 27 ◦C, 60 ◦C, and 125 ◦C. Thus, for each trojan that was run, there were 12 different tests at different temperatures and process variations. For each trojan besides the switched trojan, we tested various configurations of the trojans to determine the precision of the classifier. The switched trojan only had one configuration (on and off), while the voltage tolerance trojan was swept from 0.1 V to 0.5 V in 0.1 V increments, the parasitic capacitance trojan was tested with 10 fF, 100 fF, 1 pF, and 10 pF capacitors, and the power combiner trojan was tested with combined signals of 0.024 GHz, 0.24 GHz and 2.4 GHz. All of these data for the process technologies and temperatures were combined together for each trojan in the following way: the trojan region for the source switch trojan was determined, the same length of data for that trojan was taken for each of the trojans and then were quantized at different quantization levels (4, 6, 8, 10, 12, 14, 16, and 24) between ±0.8 to produce eight different quantized test sets, and those points were then added to an overall trojan vector test vector for each quantization level that included all the different process and temperature for that particular trojan (e.g., switch, pcap 10 f, pcomb 2.4 GHz, etc.). These vectors were then added to an overall test vector that included normal operation data and all of the variations in trojans. The training sets for all the training used 20,000 points and a test set of 1000 points. Furthermore, to determine how well the classifier can resolve individual trojans, vectors that included only normal operation with a particular trojan were included. Note that in training, the dataset included an equal number of "normal" operation and "trojan" operation sets to avoid over-training the model on trojan data.

In determining the difference between a source switch circuit, a power combiner circuit, and a parasitic capacitance, the BNN performed well over all different quantization levels. The BNN was able to determine a source switch trojan with 96% accuracy over all quantization levels, a power combiner trojan with different frequencies from 200 MHz through 2.4 GHz with nearly 100% accuracy, and an approximately 85% accuracy for the parasitic capacitance trojan over all the quantization levels. When all the different types of trojans were put into the same class and compared against the "typical" signal, there was greater than 95% testing accuracy for the BNN over the various quantization levels. Table 1 summarizes the accuracies of the different type trojans.

**Table 1.** Accuracy summary of the different type trojans.

