Location
Hilton Waikoloa Village, Hawaii
Event Website
https://hicss.hawaii.edu/
Start Date
7-1-2025 12:00 AM
End Date
10-1-2025 12:00 AM
Description
As malware becomes increasingly stealthy and more difficult to detect, behavioral malware detection has become the preferred method of detection, which uses representative run-time data from the device to determine if an infection has occurred. In this work, we collected kernel-level system calls from a router serving IoT devices during periods of benign behavior and periods of known malware infection. The system calls were processed using our custom-trained sys2vec model, which created contextual embeddings for each system call observed. We then subjected the data to a classifier using a Gated Recurrent Unit (GRU) with an Attention layer. Although this pipeline performed well for noisy, easy-to-detect malware, it struggled with stealthier malware. To combat this, we trained a classifier that uses a custom-trained BERT encoder in place of the GRU/Attention layers, which results in much better detection at a usable false positive rate (FPR) ≤ 1 × 10−5.
Recommended Citation
Carter, John; Mancoridis, Spiros; and Protopapas, Pavlos, "sysBERT: Improved Behavioral Malware Detection using BERT Trained on sys2vec Embeddings" (2025). Hawaii International Conference on System Sciences 2025 (HICSS-58). 5.
https://aisel.aisnet.org/hicss-58/st/digital_forensics/5
sysBERT: Improved Behavioral Malware Detection using BERT Trained on sys2vec Embeddings
Hilton Waikoloa Village, Hawaii
As malware becomes increasingly stealthy and more difficult to detect, behavioral malware detection has become the preferred method of detection, which uses representative run-time data from the device to determine if an infection has occurred. In this work, we collected kernel-level system calls from a router serving IoT devices during periods of benign behavior and periods of known malware infection. The system calls were processed using our custom-trained sys2vec model, which created contextual embeddings for each system call observed. We then subjected the data to a classifier using a Gated Recurrent Unit (GRU) with an Attention layer. Although this pipeline performed well for noisy, easy-to-detect malware, it struggled with stealthier malware. To combat this, we trained a classifier that uses a custom-trained BERT encoder in place of the GRU/Attention layers, which results in much better detection at a usable false positive rate (FPR) ≤ 1 × 10−5.
https://aisel.aisnet.org/hicss-58/st/digital_forensics/5