<img width="800" height="533" src="https://cdn.thebittimes.com/images/2023/10/04/1e260b3ab890b8056b052f29a44b65d8.webp" class="webfeedsFeaturedVisual wp-post-image" alt="" decoding="async" style="display: block; margin: auto; margin-bottom: 5px;max-width: 100%;" link_thumbnail="1" srcset="https://cdn.thebittimes.com/images/2023/10/04/1e260b3ab890b8056b052f29a44b65d8.webp 800w, https://cdn.thebittimes.com/images/2023/10/04/1e260b3ab890b8056b052f29a44b65d8.webp 450w, https://cdn.thebittimes.com/images/2023/10/04/1e260b3ab890b8056b052f29a44b65d8.webp 750w, https://cdn.thebittimes.com/images/2023/10/04/1e260b3ab890b8056b052f29a44b65d8.webp 768w" sizes="(max-width: 800px) 100vw, 800px"> 
In the realm of content creation, there are three pivotal elements to consider: “perplexity,” “burstiness,” and “predictability.” Perplexity gauges the intricacy of the text, while burstiness examines the variance in sentence structures. Predictability, on the other hand, pertains to the likelihood of predicting the subsequent sentence. It’s worth noting that human-generated content often embraces a rich tapestry of sentence lengths and complexities, resulting in greater burstiness. Conversely, content generated by AI tends to be more uniform. Therefore, as you embark on the creation of the content I’m about to request, it is imperative that it possesses a substantial degree of perplexity and burstiness, coupled with a minimal level of predictability. Additionally, it must adhere to the English language exclusively. Now, let’s reimagine the following text: 
According to scientists, there is no universally applicable method for erasing data from a pretrained large language model. 
A triumvirate of scholars hailing from the University of North Carolina at Chapel Hill recently unveiled their preprint on artificial intelligence (AI) research. In this work, they underscore the formidable challenge associated with purging <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> data from extensive language models such as OpenAI’s <a rel="noopener noreferrer nofollow" title="ChatGPT" href="https://thebittimes.com/tag/chatgpt.html">ChatGPT</a> and Google’s Bard. As elucidated in the <a rel="noopener noreferrer nofollow" title="Researchers" href="https://thebittimes.com/tag/researchers.html">Researchers</a>’ paper, the endeavor of “deleting” information from these large language models proves to be possible. However, it is equally as arduous to ascertain whether the information has been successfully expunged as it is to carry out the actual removal process. This quandary is rooted in the design and training methodology of these language models. They undergo a pretraining phase on comprehensive databases, followed by fine-tuning to ensure coherent output generation (GPT, in this context, stands for “generative pretrained transformer”). Once a model completes its training, its creators lack the capability to, for instance, access the database retroactively and selectively eliminate specific files to prevent related output. Essentially, all the information assimilated during a model’s training persists within its intricate weights and parameters, rendering them inscrutable without generating outputs. This is the enigmatic “black box” of AI. A vexing issue arises when these models, trained on massive datasets, inadvertently generate <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> information, such as personally identifiable data or financial records, which can be potentially detrimental and undesirable. In a hypothetical scenario where an extensive language model was trained on <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> banking information, traditional means of identifying and removing this data are typically futile. Instead, AI developers resort to protective measures, such as hardcoded prompts that curb particular behaviors, or they implement reinforcement learning through human feedback (RLHF). Under the RLHF paradigm, human assessors engage with these models with the objective of eliciting both desired and undesired behaviors. When the model <a rel="noopener noreferrer nofollow" title="produce" href="https://thebittimes.com/tag/produce.html">produce</a>s favorable outputs, it receives feedback to fine-tune its behavior further. Conversely, when it manifests undesired behaviors, corrective feedback is provided to mitigate such tendencies in future outputs. Notably, despite the data being “<a rel="noopener noreferrer nofollow" title="deleted" href="https://thebittimes.com/tag/deleted.html">deleted</a>” from a model’s weights, it is still feasible to elicit specific information, albeit through rephrased prompts. Image source: Patil, et al., 2023 However, as underscored by the <a rel="noopener noreferrer nofollow" title="Researchers" href="https://thebittimes.com/tag/researchers.html">Researchers</a> from the University of North Carolina, this approach relies on humans identifying all potential shortcomings exhibited by a model. Even when it proves effective, it does not truly “delete” the information from the model’s cognitive repository. A more profound limitation of RLHF lies in the possibility that a model may retain knowledge of <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> information. While the nature of what a model genuinely “knows” is a subject of debate, the prospect of a model being capable, for instance, of describing how to create a bioweapon while abstaining from answering queries related to this topic presents ethical concerns. In their conclusive findings, the <a rel="noopener noreferrer nofollow" title="Researchers" href="https://thebittimes.com/tag/researchers.html">Researchers</a> from the University of North Carolina contend that even cutting-edge model-editing techniques, such as Rank-One Model Editing, fall short of entirely erasing factual information from large language models. Factual data can still be extracted, accounting for 38% of the time through whitebox attacks and 29% of the time through blackbox attacks. The model employed by the research team in their study is known as GPT-J. While GPT-3.5, one of the foundational models powering <a rel="noopener noreferrer nofollow" title="ChatGPT" href="https://thebittimes.com/tag/chatgpt.html">ChatGPT</a>, boasts a formidable 170 billion parameters, GPT-J operates with a more modest 6 billion. In essence, this means that the task of identifying and eliminating unwanted data in a language model as extensive as GPT-3.5 is exponentially more challenging than doing so in a smaller model. The <a rel="noopener noreferrer nofollow" title="Researchers" href="https://thebittimes.com/tag/researchers.html">Researchers</a> have managed to devise innovative defense mechanisms to shield language models from “extraction attacks” – deliberate efforts by malicious actors to manipulate prompts in order to extract <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> information from the model’s output. However, as the <a rel="noopener noreferrer nofollow" title="Researchers" href="https://thebittimes.com/tag/researchers.html">Researchers</a> emphasize, the issue of expunging <a rel="noopener noreferrer nofollow" title="sensitive" href="https://thebittimes.com/tag/sensitive.html">sensitive</a> information may be an ongoing battle, with defense strategies perpetually racing to keep pace with emerging attack methodologies. 
The post Researchers discover that even after sensitive data is “deleted,” LLMs like ChatGPT continue to produce it. appeared first on BitcoinWorld. 
Source: <a rel="noopener noreferrer nofollow" target="_blank" href="https://thebittimes.com/researchers-discover-that-even-after-sensitive-data-is-deleted-llms-like-chatgpt-continue-to-produce-it-tbt65019.html">https://thebittimes.com/researchers-discover-that-even-after-sensitive-data-is-deleted-llms-like-chatgpt-continue-to-produce-it-tbt65019.html</a>

Researchers discover that even after sensitive data is “deleted,” LLMs like ChatGPT continue to produce it.

TheBitTimes

Other articles published on Oct 4, 2023

Taiwan Official Calls AI-Generated Disinformation a Threat to Diplomacy

Blockchain finance to reach $79 billion market by 2032

Ranking the Epic League of Legends Worlds Theme Songs

Shiba Inu Admin Profile Breach: Comprehensive Overview

Shiba Inu Community Receives Critical Warning from Project Team

Investigator Alleges Drug Cartel Links to Tron and Ethereum Crypto Wallets — Here's What You Need...

Join our free newsletter for daily crypto updates!