Preface : After investigating so many directions, I have some insights into the emergence and development of research directions, especially in the field of AI security. In my opinion, the so-called valuable and meaningful research refers to research that can make waves in real society and produce certain benefits for the country and society.
For example, in the field of AI security, in 2016, Alpha Go defeated world Go champion Lee Sedol, and the development of neural networks ushered in a wave. After that, the security issue of neural networks began to attract much attention, and some research directions emerged as the times require -
-
On January 15, 2017, the first copyright protection work for neural network models was posted on arXiv, and was later published on ICMR in April of the same year:
-
On August 22, 2017, the first backdoor attack work on neural network models was posted on arXiv, and was later published in IEEE Access in April 2019:
-
In 2017, the first generative steganography work based on language model was published in the ACL Student Research Workshop:
In 2022, ChatGPT was born, and from January 2023, several works were posted on arXiv.
- In June 2023, the first article on copyright protection for large models was published at ICML 2023.
Looking at the development history of academic research, it is not difficult to see that the birth of any research direction is not accidental. Corresponding research is often carried out because it has certain commercial or social value, or makes certain contributions to the country.
In addition, sensational events that affect a wide range of areas, especially world-class events, can often promote the birth of a certain research field. For example, after the research direction of neural network copyright protection was proposed, most research work was carried out in the field of image classification such as face recognition, and little consideration was given to the field of natural language generation. However, in the large language model represented by ChatGPT After its emergence, watermark research on generative language models began to enter the field of vision of scholars.
Therefore, sometimes it is important to have an overall view of thinking about the future. It can predict the development path of scientific research to a certain extent, and thus make corresponding adjustments to one's own research direction. While we are witnessing history, we can also try to change the trajectory of history.
Postscript : The above insights were generated during the investigation of "large model watermarks". Our instructors have always taught us to be leaders, not followers, so this article encourages you to reach a higher level on the road to academic research!