METHODS FOR DETECTION AND COUNTERACTION OF CYBERATTACKS OF THE TYPE GPS SPOOFING AND GPS JAMMING USING AI FOR DIFFERENTIAL CORRECTION SYSTEMS AND GLOBAL NAVIGATION SATELLITE SYSTEM

Abstract

 An integrated approach combining clustering, signature analysis, and forecasting methods is proposed to effectively detect and neutralize cyberattacks in real time. 
 The purpose of the article is to develop methods for using artificial intelligence (AI) for cyber defense of Differential Global Positioning Systems (DGPSs) of global navigation satellite systems (GNSS), autonomous autonomous DGPS, against GPS spoofing and GPS jamming attacks. 
 Scientific novelty. For the first time, a hybrid methodology based on K-Means, Fuzzy C-Means, Online K-Means, Random Forest, and Recurrent Neural Network (RNN) algorithms using GRU layers is proposed. This technique allows not only to detect anomalies but also to adaptively update the model in response to new threats. A unique GNSS dataset was synthesized, including 500 samples for each class (normal signals, spoofing, jamming) with 8 critical features (signal strength, delay, acceleration, etc.). For the first time, post-quantum cryptography mechanisms are integrated to improve the system security. 
 Results. The study conducted a clustering analysis using K-Means and Fuzzy C-Means, which detected 66.7% of anomalies (100 out of 150 expected), while online K-Means showed better adaptability, detecting 204 anomalies. The distribution of clusters for K-Means was [100, 154, 1246], for Fuzzy C-Means - [1246, 154, 100], and for online K-Means - [924, 204, 372]. Signature analysis using Random Forest effectively filtered out false signals, the model was trained on 70% of the data and tested on 30%. Predicting attacks using GRU-based RNNs achieved an accuracy of 0.50 on the training set and 0.51 on the validation set, using 3 GRU layers of 128 neurons each, training was stopped after 32 epochs. Countermeasures included adaptive synchronization with a time correction of 0.14 ms, switching to inertial navigation when more than 50% of anomalies were detected, and simulating post-quantum key generation. Additional analyses revealed a correlation between power and signal latency of 0.54, and multisensory integration used an acceleration threshold of 0.5. Nine PNG graphs were generated to visualize various aspects of the analysis. Data processing included the use of 500 samples for each class (normal, spoofing, jamming) with 8 features, and data standardization for RNNs was applied. 
 Conclusions. The effectiveness of the proposed hybrid approach for protecting autonomous SDKs has been proved. Online K-Means showed the best results due to its ability to adapt to dynamic conditions. Further research should focus on using real GNSS data; improving the RNN architecture; expanding the feature set to improve accuracy.