Real-time economic data analysis for false data embedding detection

Detection of data injection is a significant area of research due to the vulnerability of systems to cyberattacks that manipulate state estimation by injecting deceptive data into sensor readings, evading security system detection. Such manipulations can cause state estimation outputs to deviate from their true values. This study proposes a methodology for detecting false data injection using five deep neural network models. This approach utilizes sliding windows to generate online error vectors, enabling the identification and classification of malicious data within sensor measurements. To assess the effectiveness of the proposed methodology, this approach is implemented on the daily closing prices of the S&P 500 Index from 2013 to 2024, with false data injected via software that maintains statistical characteristics. The results demonstrate that the convolutional neural network delivers the most favorable performance, achieving over 94% accuracy and an F1-score of 0.95. These results highlight the strong potential of deep neural networks as effective tools for detecting false data in measurements obtained from diverse sources. • An online, model-free framework for detecting false data injections in economic time series is proposed in this work. • The injected false data are designed to remain within normal statistical ranges, making the detection problem substantially more challenging. • A residual-based error vector representation is introduced to enhance anomaly signals and improve neural classification performance. • Five deep neural network architectures (MLP, CNN, LSTM, LSTM-FCN, and CAE) are comparatively evaluated under the same experimental conditions. • The proposed approach is validated on real-world S&P 500 Index data (2013 – 2024), achieving accuracy above 94% and demonstrating the CNN’s superiority for practical online anomaly detection.