GPT를 활용한 MQTT 기반 IDS 우회 방법 = MQTT-based IDS Evasion Method using GPT

국문 초록 (Abstract)

사물인터넷(IoT)은 경량 프로토콜인 MQTT(Message Queuing Telemetry Transport)를 통해 기기 간의 원활한 통신을 지원하며, 효율성과 확장성 등의 장점으로 인해 다양한 분야에서 활용되고 있다. 그러나 IoT가 확산되며 네트워크 보안 문제도 야기되고 있으며, 특히 MQTT 트래픽의 경량성과 비표준화된 특성, 그리고 공격 기법의 지속적인 진화로 인해 기존 침입 탐지 시스템(IDS)은 이러한 트래픽을 효과적으로 탐지하는 데 한계를 드러내고 있다. 본 연구에서는 IDS의 탐지 성능을 향상시키는데 활용할 수 있도록 GPT(Generative Pre-trained Transformers) 같은 생성 인공지능(AI)를 활용하여 IDS 탐지를 회피할 수 있는 MQTT 데이터를 생성하는 방법을 제안한다. 전처리된 정상 MQTT 데이터를 기반으로 GPT 모델을 학습시켜 데이터를 생성하고, 악성 MQTT 데이터를 merge 함으로써 IDS 탐지를 우회할 수 있는 데이터를 생성하였다. 생성된 데이터는 IDS의 취약점을 분석하고, 탐지 성능 개선을 위한 학습 데이터 강화 및 모델 설계에 활용될 것을 기대한다.

다국어 초록 (Multilingual Abstract)

The Internet of Things(IoT) supports seamless communication between devices through the lightweight protocol Message Queuing Telemetry Transport(MQTT), which is widely applied across various fields due to its advantages in efficiency and scalability. However, the rapid expansion of IoT has raised significant network security concerns. In particular, the lightweight and non-standardized characteristics of MQTT traffic, coupled with the continuous evolution of attack techniques, have exposed the limitations of existing Intrusion Detection Systems(IDS) in effectively detecting such traffic. This study proposes a method for generating MQTT data capable of evading IDS detection by leveraging Generative AI, specifically Generative Pre-trained Transformers(GPT), to enhance IDS detection performance. By training a GPT model on preprocessed normal MQTT data, malicious MQTT data were merged and synthesized to create datasets capable of bypassing IDS detection. The generated data are expected to facilitate the analysis of IDS vulnerabilities and contribute to improving detection performance through the enhancement of training data and model design.

참고문헌 (Reference)

1 "Security threats caused by AI, including deepfake crimes"

2 "SandsLab technology exports"

3 "Rising AI hacking threats: Criminal AIs in 2024"

4 S. Choi, "Novel Feature Extraction Method for Detecting Malicious MQTT Traffic Using Seq2Seq" 12 (12): 12306-, 2022

5 이든 ; 임승순 ; 최선오, "Malicious traffic detection method using LSTM and sliding window in MQTT-based IoT environment" 21 (21): 111-120, 2023

6 S. O. Choi, "Malicious PowerShell Detection using Attention against Adversarial Attacks" 9 (9): 1817-, 2020

7 A. Radford, "Language models are unsupervised multitask learners" 1 (1): 2019

8 "IoT vulnerability"

9 "IoT security risks"

10 N. Butt, "Intelligent Deep Learning for Anomaly-Based Intrusion Detection in IoT Smart Home Networks" 10 (10): 4598-, 2022

11 B. Isong, "Insights into Modern Intrusion Detection Strategies for Internet of Things Ecosystems" 13 (13): 2370-, 2024

12 I. J. Goodfellow, "Generative Adversarial Networks" 2014

13 OpenAI, "GPT-4 introduction"

14 T. Fladby, "Evading a Machine Learning-based Intrusion Detection System through Adversarial Perturbations" 161-166, 2020

15 신상윤 ; 이다희 ; 이상진, "Design Method of Things Malware Detection System(TMDS)" 33 (33): 459-469, 2023

16 "Deepfakes in Russo-Ukrainian information warfare"

17 "Cybersecurity threats in 2024: Hacktivism and AI"

18 전용희, "A Study on the Security Modeling of Internet of Things(IoT)" 15 (15): 17-27, 2017