Rank-based Modeling for Universal Packets Compression in Multi-modal Communications

This repository contains partial code for the ByteTrans framework.

ByteTrans Framework

An Example of Lossless Packets Data Compression

The predictive model outputs a probability distribution over 256 possible byte values (0–255), with probabilities summing to 1. Each byte is assigned a unique rank based on its probability, where rank 0 corresponds to the most probable byte, rank 1 to the second most probable byte, and so on. This ensures every byte has a deterministic and unique rank.

Suppose the packet sequence needed to be compressed is [120, 85, 90], the current input byte is 120. The model outputs the following probabilities for the next byte:

Byte Value	Probability	Rank
85	0.40	0
67	0.30	1
90	0.20	2
...	...	...

Here, the byte 85 is the most probable, so it is assigned rank 0. During compression, only the rank (0) is stored instead of the actual byte value (85).

Now consider predicting the third byte, given the previous two bytes [120, 85]. The model outputs:

Byte Value	Probability	Rank
67	0.35	0
90	0.25	1
50	0.15	2
...	...	...

The true value of the third byte is 90, which corresponds to rank 1 based on the model’s predictions. I think this is what you called a “prediction error.” However, we do not store rank 0. Instead, the rank 1 is stored, representing the byte 90.

---

During decompression, the stored ranks are retrieved and mapped back to the original byte values using the same predictive model. Here’s how the process works for the given example:

Given the input byte 120 and the stored rank 0, the predictive model outputs the following probabilities:

Byte Value	Probability	Rank
85	0.40	0
67	0.30	1
90	0.20	2
...	...	...

Since the stored rank is 0, the model retrieves the byte 85 as the most probable value.

Given the first two bytes [120, 85] and the stored rank 1, the predictive model outputs the following probabilities:

Byte Value	Probability	Rank
67	0.35	0
90	0.25	1
50	0.15	2
...	...	...

Since the stored rank is 1, the model retrieves the byte 90, which corresponds to the correct value.

---

By storing the ranks (0 for 85 and 1 for 90), the model ensures that the original packet sequence [120, 85, 90] is perfectly reconstructed. This deterministic mapping guarantees lossless compression, as the same predictive model is used during both compression and decompression to consistently interpret the ranks.

Citation

If you want to cite our paper in your work, please use the following BibTeX entry.

@inproceedings{xuanhaoluoByteTrans2025,
  title = {{Rank-Based Modeling for Universal Packets Compression in Multi-Modal Communications}},
  author = {Luo, Xuanhao and Peng, Zhiyuan and Li, Zhouyu and Yu, Ruozhou and Liu, Yuchen},
  booktitle = {{IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM)}},
  year = {2025},
}