TSCL:Multi-party loss Balancing scheme for deep learning Image steganography based on Curriculum learning

Abstract: For deep learning-based image steganography frameworks, in order to ensure the invisibility and recoverability of the information embedding, the loss function usually contains several losses such as embedding loss, recovery loss and steganalysis loss. In previous research works, fixed loss weights are usually chosen for training optimization, and this setting is not linked to the importance of the steganography task itself and the training process. In this paper, we propose a Two-stage Curriculum Learning loss scheduler (TSCL) for balancing multinomial losses in deep learning image steganography algorithms. TSCL consists of two phases: a priori curriculum control and loss dynamics control. The first phase firstly focuses the model on learning the information embedding of the original image by controlling the loss weights in the multi-party adversarial training; secondly, it makes the model shift its learning focus to improving the decoding accuracy; and finally, it makes the model learn to generate a steganographic image that is resistant to steganalysis. In the second stage, the learning speed of each training task is evaluated by calculating the loss drop of the before and after iteration rounds to balance the learning of each task. Experimental results on three large public datasets, ALASKA2, VOC2012 and ImageNet, show that the proposed TSCL strategy improves the quality of steganography, decoding accuracy and security.

• The paper introduces a two-stage curriculum learning framework that dynamically adjusts loss weights for embedding, recovery, and steganalysis tasks.
• It employs both discrete and continuous scheduling methods to shift focus during training based on predefined curricula and loss dynamics.
• Experimental results on datasets like ALASKA2, VOC2012, and ImageNet demonstrate improved imperceptibility and decoding accuracy over fixed-weight schemes.

Deep learning-based image steganography frameworks often employ multiple loss functions during training to ensure the invisibility of the hidden message, the accuracy of message recovery, and resistance to steganalysis. These losses typically include an embedding loss (measuring the difference between the cover image and the stego image), a recovery loss (measuring the difference between the original and recovered secret message), and a steganalysis loss (adversarially training against a steganalyst to make the stego image undetectable).

A common approach is to combine these losses using fixed weights. However, this can be suboptimal because the training dynamics of different tasks (embedding, recovery, steganalysis) vary throughout the training process. The paper "TSCL: Multi-party loss Balancing scheme for deep learning Image steganography based on Curriculum learning" (2504.18348) addresses this by proposing a Two-stage Curriculum Learning loss scheduler (TSCL) to dynamically balance these multiple losses.

The core idea behind TSCL is to adapt the loss weights based on both a predefined curriculum (task importance) and the observed training progress (loss dynamics). This acknowledges that different aspects of the steganography task become more or less important at different stages of training, and that tasks may learn at different speeds.

The TSCL scheme consists of two distinct phases:

1. A Priori Curriculum Control Stage:
   • This stage is based on the intuition that in image steganography, imperceptibility is paramount, followed by recovery accuracy, and then security against steganalysis.
   • The training is divided into periods where the model's focus is shifted by adjusting the loss weights.
   • Initially, the embedding loss is prioritized to ensure the stego image is visually similar to the cover image.
   • Then, the decoding loss is prioritized to improve the accuracy of secret message recovery.
   • Finally, the steganalysis loss is prioritized to make the stego image harder for a steganalyst to detect.

   • The paper explores both discrete and continuous methods for adjusting these weights across epochs.

     • Discrete Scheduling: Weights are changed abruptly at fixed iteration steps. Different strategies involve fixing or unfixing the iteration step size and the magnitude of the weight adjustment.

     • Continuous Scheduling: Weights are changed gradually according to a mathematical function of the training epoch. The paper tests linear, sinusoidal, and exponential functions to control the rate of weight increase for a given loss during its prioritized phase. For instance, a linear function would linearly increase a loss weight $\alpha$ from an initial value $a_0$ to a final value $a_2$ between epochs $C_1$ and $C_2$:

       $$\alpha = \begin{cases} a_0, & \text{epoch} < C_1 \ a_0 + (a_2 - a_0) \times \frac{\text{epoch}-C_1}{C_2-C_1}, & C_1 \leq \text{epoch} < C_2 \ a_2, & \text{epoch} \geq C_2 \end{cases}$$

* This stage ensures the model builds capabilities layer by layer, focusing on fundamental requirements before refining performance on more complex adversarial objectives.

1. Loss Dynamics Control Stage:
   • After the initial curriculum stage, the training enters this phase where loss weights are adjusted dynamically based on how quickly each loss is decreasing.
   • The core idea is that if a task's loss is decreasing slowly, it indicates that the task is currently difficult for the model to learn, and its corresponding weight should be increased to give it more training focus in the next iteration.
   • The learning speed of a task $k$ at iteration $t$ is measured by the ratio of the loss at iteration $t-1$ to the loss at iteration $t-2$: $a_0$0. A value close to 1 means slow progress, while a smaller value means faster progress.
   • These dynamic ratios are then multiplied by a set of predefined a priori coefficients ($a_0$1) to maintain the relative importance established in the first stage. The weight for task $a_0$2 at iteration $a_0$3 becomes $a_0$4. The paper sets $a_0$5 (specifically 1, 0.8, 0.4 in experiments) to reflect the overall importance hierarchy.
   • This stage provides an adaptive mechanism to balance the learning of different tasks based on their current optimization difficulty, preventing simple tasks from overfitting while difficult tasks lag behind.

Practical Implementation:

Implementing TSCL involves integrating this dynamic weight calculation logic into the training loop of a deep learning steganography model. Assuming a standard adversarial training setup with an encoder, decoder, and steganalyst, the training loop structure would involve:

• Calculating the three primary losses: encoding loss ($a_0$6), decoding loss ($a_0$7), and steganalysis loss ($a_0$8). Note that $a_0$9 is used differently for updating the encoder/decoder (minimize detection likelihood) and the steganalyst (maximize detection likelihood).
• At each training iteration (or epoch, as described in the paper's evaluation), determining the current epoch number.
• Based on whether the current epoch falls within the A Priori Curriculum Control stage or the Loss Dynamics Control stage, calculating the weights for each loss:
  • If in Stage 1: Use the predefined schedule (discrete steps or continuous function) to get the weights.
  • If in Stage 2: Calculate the loss ratio for each task based on the losses from the two previous iterations/epochs and multiply by the a priori coefficients.
• Combine the losses using the calculated weights: $a_2$0. (The steganalyst has its own objective, typically minimizing $a_2$1 with respect to its own parameters, using samples labeled as 'stego' and 'cover').
• Perform backpropagation and optimizer steps using the weighted total loss for the encoder/decoder and the steganalyst's loss for the steganalyst.
• Crucially, maintain a history of the individual loss values (e.g., averaged per epoch) to calculate the loss ratios needed for Stage 2.

Here is simplified pseudocode demonstrating the weight calculation logic within the training loop:

$a_2$4

Implementation Considerations and Trade-offs:

• Hyperparameter Tuning: The performance of TSCL is sensitive to the choice of curriculum schedule parameters ($a_2$2, function types for continuous, step sizes/magnitudes for discrete) and the a priori coefficients ($a_2$3). These need to be tuned based on the specific model architecture, dataset, and desired performance characteristics (balancing imperceptibility, accuracy, and security).
• Computational Overhead: The computational cost of calculating loss ratios and updating weights is minimal compared to the cost of forward and backward passes through the neural networks.
• Choosing Curriculum Functions/Schemes: The experiments show that the continuous sinusoidal function improves PSNR but might decrease accuracy, while discrete schemes can improve both depending on settings. The "unfixed iteration step size and weight adjustment amplitude" discrete scheme performed well in the paper's tests. Practitioners may need to experiment with different schedules.
• Loss History: Maintaining and accessing loss history adds a small memory overhead but is necessary for Stage 2. The paper calculates ratios based on epoch-averaged losses, simplifying the history requirement compared to per-batch tracking.
• Generalizability: The TSCL principle (prioritizing based on curriculum and adapting based on dynamics) can be applied to other multi-task learning or multi-loss optimization problems beyond steganography where task importance and learning difficulty change over time.
• Task Definition: Clearly defining the loss functions and how they contribute to the overall objectives (imperceptibility, recovery, security) is crucial for setting up the curriculum and a priori priors effectively.

The experimental results on ALASKA2, VOC2012, and ImageNet datasets show that TSCL generally improves performance across imperceptibility (SSIM, MSSSIM, PSNR, RMSE) and decoding accuracy compared to a fixed-weight baseline. Security against steganalysis showed mixed results, improving on ALASKA2 but slightly decreasing on ImageNet, highlighting the inherent trade-offs in steganography tasks and the challenge of balancing all objectives perfectly. The visual results further support that TSCL leads to stego images with fewer noticeable artifacts. The paper demonstrates that the two-stage approach is more effective than using either stage in isolation, confirming the synergy between structured curriculum learning and dynamic adaptation.