Detection and Obfuscation of Deepfake Texts

Adaku Uchendu

UNCLASSIFIED

UNCLASSIFIED

RAAINS - 1

Author Initials 11/13-15/19

Outline

1. Introduction
2. Motivation
3. Problem Definition
1. Authorship Attribution (AA)
2. Authorship Obfuscation (AO)
4. Landscape of AA & AO
5. Open Problems
6. Conclusion & Future Work

​

​

​

RAAINS - 2

Author Initials 11/13-15/19

Introduction – Turing Test

• Turing Test: is an intelligence test, where a human decides if they are speaking to a human or machine.

​

​

Uchendu, A., Cao, J., Wang, Q., Luo, B., & Lee, D. (2019). Characterizing Man-made vs. Machine-made Chatbot Dialogs. In TTO.

RAAINS - 3

Author Initials 11/13-15/19

Introduction – Reverse Turing Test

• Turing Test: is an intelligence test, where a human decides if they are speaking to a human or machine.

​

​

​

• Reverse Turing Test: is an intelligence test, where a Computer Algorithm decides if they are speaking to a human or machine. This is Automatic Turing Test.

​

​

Uchendu, A., Cao, J., Wang, Q., Luo, B., & Lee, D. (2019). Characterizing Man-made vs. Machine-made Chatbot Dialogs. In TTO.

RAAINS - 4

Author Initials 11/13-15/19

Motivation – Deepfake

RAAINS - 5

Author Initials 11/13-15/19

What are Neural Text Generators (NTGs)?

RAAINS - 6

Author Initials 11/13-15/19

Motivation – Evolution of Neural Text Generators 

• NTGs are becoming ubiquitous at an alarming rate 
• Newer NTGs are more sophisticated 
• Maliciously used to generate misinformation
• Huggingface's data repo currently houses more than 3500 English Text generators

Uchendu, A., Ma, Z., Le, T., Zhang, R., & Lee, D. (2021, November). TURINGBENCH: A Benchmark Environment for Turing Test in the Age of Neural Text Generation. In Findings of the Association for Computational Linguistics: EMNLP 2021 (pp. 2001-2016).

RAAINS - 7

Author Initials 11/13-15/19

Authorship Attribution (AA)​

• AA is the task of accurately assigning a given document to its true author ​

​

Uchendu, A., Le, T., Shu, K., & Lee, D. (2020, January). Authorship attribution for neural text generation. In Conf. on Empirical Methods in Natural Language Processing (EMNLP).

RAAINS - 8

Author Initials 11/13-15/19

Authorship Obfuscation (AO)

• AO is the task of masking an author's writing style/signature to conceal identity, usually for privacy reasons (ex: activists)​

​

Mahmood, A., Ahmad, F., Shafiq, Z., Srinivasan, P., & Zaffar, F. (2019). A Girl Has No Name: Automated Authorship Obfuscation using Mutant-X. Proc. Priv. Enhancing Technol., 2019(4), 54-71.

RAAINS - 9

Author Initials 11/13-15/19

Why does AA & AO matter for neural text detection?

Uchendu, A., Le, T., & Lee, D. (2022). Attribution and Obfuscation of Neural Text Authorship: A Data Mining Perspective. arXiv preprint arXiv:2210.10488.

RAAINS - 10

Author Initials 11/13-15/19

Data Generation Process

• Collect human-written documents​
• Use Title of human-written document as prompt​
• With Prompt, generate texts with NTGs​
• NTG texts is similar to human-written​

​

Uchendu, A., Le, T., Shu, K., & Lee, D. (2020, January). Authorship attribution for neural text generation. In Conf. on Empirical Methods in Natural Language Processing (EMNLP).

RAAINS - 11

Author Initials 11/13-15/19

Authorship Attribution for Neural Text 

• Given a text T1, which NTG (among k options) generated T1 (k ≥ 2)?

​

​

​

​

​

• Most researchers study the AA problem, where k=2 authors – human & machine/neural method

​

​

Uchendu, A., Le, T., Shu, K., & Lee, D. (2020, January). Authorship attribution for neural text generation. In Conf. on Empirical Methods in Natural Language Processing (EMNLP).

RAAINS - 12

Author Initials 11/13-15/19

Categories of AA solutions for neural text detection

Uchendu, A., Le, T., & Lee, D. (2022). Attribution and Obfuscation of Neural Text Authorship: A Data Mining Perspective. arXiv preprint arXiv:2210.10488.

RAAINS - 13

Author Initials 11/13-15/19

Categories of AO techniques used to obfuscate neural text authorship

Uchendu, A., Le, T., & Lee, D. (2022). Attribution and Obfuscation of Neural Text Authorship: A Data Mining Perspective. arXiv preprint arXiv:2210.10488.

RAAINS - 14

Author Initials 11/13-15/19

Open Problems

Uchendu, A., Le, T., & Lee, D. (2022). Attribution and Obfuscation of Neural Text Authorship: A Data Mining Perspective. arXiv preprint arXiv:2210.10488.

RAAINS - 15

Author Initials 11/13-15/19

Conclusion

• Deepfake/Neural Text Detection (NTD) is non-trivial
• AA solutions for NTD are susceptible to AO techniques
• DL-based (Transformers) are more accurate than Stylometric & Statistical
• Stylometric & Statistical are more adversarially robust to AO techniques than DL-based (Transformers)
• Hybrid AA models are the best (robustness, accuracy & generalizability)
• Human performance in detection of neural texts is at chance-level
• Need better training techniques to improve human performance

​

​

PCA Visualization of BERT learned weights​

F1 score – 80%

​

1. Uchendu, A., Ma, Z., Le, T., Zhang, R., & Lee, D. (2021, November). TURINGBENCH: A Benchmark Environment for Turing Test in the Age of Neural Text Generation. In Findings of the Association for Computational Linguistics: EMNLP 2021 (pp. 2001-2016).
2. Uchendu, A., Le, T., & Lee, D. (2022). Attribution and Obfuscation of Neural Text Authorship: A Data Mining Perspective. arXiv preprint arXiv:2210.10488.

​

​

RAAINS - 16

Author Initials 11/13-15/19

Future Work

RAAINS - 17

Author Initials 11/13-15/19

About Me

• Name: Adaku Uchendu (she/her)
• NSF Scholarship for Service Scholar
• Alfred P. Sloan Scholar
• Email: azu5030@psu.edu
• Affiliation: PIKE Lab, Advisor: Dr. Dongwon Lee
• School: 
• 5th-Year Ph.D. student at The Pennsylvania State University
• B.S. Mathematics, UMBC (University of Maryland Baltimore County)
• Research Focus: NLP, specifically NLG and Machine learning
• Personal Website: https://adauchendu.github.io/

​

RAAINS - 18

Author Initials 11/13-15/19