Abstract
This research aimed to investigate and implement an image steganography technique that does not interfere with an existing image recognition artificial neural network system. Many systems today use neural networks in their day-to-day operations. Hence there arises a need to study information-hiding techniques such as steganography and investigate their impact on artificial neural network systems. This study contributes to advancements in integrating secure information-hiding with digital processes, including the widely used simulated intelligence methods.
Three image steganography methods were tested: traditional LSB, proposed DCT LSB-2, and a novel DCT LSB-2 edge encoding technique. The traditional LSB method was adopted from the literature survey. This method is crucial to validate if the results obtained from the proposed and novel methods align with the theory and expected output. The proposed DCT LSB-2 method concatenates two techniques: Discrete Cosine Transform and Least Significant Bit 2. The proposed method was further enhanced to form a novel DCT LSB-2 edge encoding technique. The novel approach works very similarly to the proposed method, except that the novel method encodes data bits along a specific pixel surface area of a defined size and not in all the cover image’s pixels.
The methods tested in this work allow data to be hidden in a cover image pixel. In contrast, the image recognition neural network checks whether the stego-image is still recognizable or not. The resulting stego-images from the three methods were analyzed using an image recognition neural network implemented in the Keras TensorFlow soft tool. The results showed that using the proposed DCT LSB-2 encoding technique allows a high data payload and minimizes visible alterations, hence maintaining the efficiency of the neural network compared to the traditional LSB method. From the results, an optimum ratio for encoding data in an image was determined to maintain the high robustness of the steganography system. The proposed DCT LSB-2 technique has improved stego-system performance compared to the previously tested traditional LSB methods obtained from the literature. To further enhance the efficiency of the neural network, the novel DCT LSB-2 edge encoding technique was introduced and has further demonstrated improved performance compared to the proposed DCT LSB-2 method.
The findings show that the proposed novel DCT LSB-2 method enables the functionality of information hiding in a cover image while simultaneously maintaining the prediction efficiency of an image recognition artificial neural network.
Keywords: Least Significant Bit; Discrete Cosine Transform; Artificial Neural Network; Peak-Signal-to-Noise Ratio; Percentage Bytes Changed; Bitmap; Artificial Intelligence.