Quaid Khalid’s Post

DPAdapter: A New Technique Designed to Amplify the Model Performance of Differentially Private Machine Learning DPML Algorithms by Enhancing Parameter Robustness Privacy in machine learning is critical, especially when models are trained on sensitive data. Differential privacy

Please join the Birmingham (AL) IAPP KnowledgeNet 15 March 2024 to hear Katharine Jarmul speak on Large AI Models and Their Risk to Privacy. Topic: Large AI Models and Their Risk to Privacy Join the Birmingham (AL) KnowledgeNet Chapter for a presentation that will cover exactly how large AI models create unique threats to individuals and their data privacy. Told from the perspective of a privacy technologist and machine learning engineer, this talk will aim to upskill technical understanding of such systems and may change notions of how to effectively build and use AI systems. Attendees will learn the latest in both privacy research and observed behavior of such models, and potential approaches to combat such privacy issues, some of which may be surprising! Speaker: Katharine Jarmul, Principal Data Scientist, Thoughtworks * Please check out Katharine's wonderful book - Practical Data Privacy https://lnkd.in/eR437cxD Time: 12:00-13:00 CDT Eligible CPE Credits: CIPP/A, CIPP/C, CIPP/E, CIPP/US, CIPM, and CIPT. 1.0 CPE In-Person Registration: https://lnkd.in/ek4PP3U7 *Thank you to Burr & Forman LLP for providing food and beverages for the In-Person event! Livestreamed Registration: https://lnkd.in/daPED2hD

🌐 Safeguarding Privacy in Machine Learning 🤖 In our latest post, we discussed how differential privacy can protect sensitive information in training data without compromising utility using methods like Differential-private Synthetic Data and DP-SGD. Read more: https://lnkd.in/eRBeKHbq This article is a short primer on one such technique we use for privacy in ML. It is aimed to be part of a series of articles on Privacy and Machine Learning....and how we are using engineering+ research to scale PETs for ML :) #Privacy #DataProvenance #TrainingData #DataSecurity #DifferentialPrivacy #ValyuExchange Harvey Yorke

OpenAI's DEATH BLOW to Open Source 📜 Overview: OpenAI recently made public statements about AI principles and potential regulation, sparking discussions and concerns particularly around the impact on open source AI. 🔒 Security and Privacy: Emphasized the importance of protecting AI model weights, considering them critical to preserving the integrity and value of AI technologies. 🛡️ Proposed Measures: Suggested advanced security protocols for AI accelerators, including trusted computing and encrypted model weights to enhance protection against unauthorized access. 📊 Impact on Open Source: The approach could restrict open source AI development by necessitating proprietary, high-security hardware for running advanced models, potentially centralizing control over AI technologies. 🏢 Hardware and Compliance: Discussed the need for robust physical security at data centers and compliance standards tailored for AI, to guard against both internal and external threats. 🔐 Cyber Defense Potential: Highlighted AI's transformative potential in cyber defense, capable of significantly improving security automation and response capabilities. #OpenAI #AISecurity #CyberDefense #OpenSourceAI #AIRegulation #ModelProtection #DataPrivacy #AIAccelerators #TechInnovation #AICompliance

💻 scrubadub_spacy: 38⭐ Protect sensitive data with ease 🔒👌 Scrubadub_spacy makes cleaning personally identifiable information a breeze. Safeguard your data and comply with privacy regulations by using this powerful open-source tool to automatically identify and remove sensitive information from unstructured text. The links are as always a side-quest: Check it out here: https://lnkd.in/e-csycCb My newsletter subscribers learned about this 25 months ago! https://late.email ┈┈┈┈┈┈┈┈✁┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈ 👋 Moin, I'm Jesper! I share non-hype AI like this every day to help you build better real-world ML applications! 𝗙𝗼𝗹𝗹𝗼𝘄 Jesper Dramsch to stay in the loop! Join 3,000 others here: https://lnkd.in/gW_-ym7A #DataScience #Programming #ArtificialIntelligence #Data #LateToTheParty #Career #DeepLearning

After a week in New York, #legalweek2024 has now come to an end. The insights gained and connections made are invaluable for navigating the ever-evolving legal landscape. As legal professionals from all over the world gathered to discuss the latest trends and advancements in the industry, it was clear that the legal landscape is in the midst of a rapid transformation. From the rise of generative artificial intelligence and machine learning to the increasing importance of cybersecurity, there were many topics that were at the forefront of the discussion. One of the key takeaways from #legalweek2024 was the importance of staying agile and adaptable in the face of change. With so many new technologies and innovations emerging all the time, it's essential for legal professionals to stay up-to-date and be willing to pivot as needed to stay ahead of the curve. #triagedata #tds #generativeai #legalweek #legaltech

Did you know that today is National AI Literacy today, per a number of groups focused on AI in schools? Well just in time - Our Youth/Ed will have a guide for schools out shortly on Generative AI in schools - data protection policy brief. Join me - now - to discuss this issue with FPF's David Sallay Future of Privacy Forum On my page NOON ET sharp.

Durk Kingma, co-founder of OpenAI, joins AnthropicShine light on data-centric tasks and their intricacies🔎 Our latest article explores the complexities around requests that cannot be processed due to the nature of the task. Specifically, the piece underscores the importance of privacy in the operations of data financing and the potential implications it can have if mishandled. 💼🌐 The discussion seeks to equip you with the necessary knowledge on how to navigate these sensitive processes while maintaining respect for personal data. ➡️ Browse through it here: https://lnkd.in/d_EqUmNb Join the conversation as we delve deeper into this crucial subject! 💡📚 #DataProcessing #Privacy #OperationsManagement #DataAnalytics [Image associated with the article]

It's my pleasure to attend PETS this year. I had a great time with many experts in the privacy and security community, talking about potential privacy breaches and remedies in LLM and federated learning. I also presented our recent work on FedLAP-DP: Federated Learning by Sharing Differentially Private Loss Approximations. It is an interesting work that overcomes non-IID + DP optimization in FL by approximating local loss landscapes and communicating synthetic images (instead of gradients!). Such a novel optimization strategy provides a new option to FL users, especially when considering extremely non-IID and DP scenarios. Please take a look if you are interested in more details. [paper] https://lnkd.in/eFRAzCs4 [code] https://lnkd.in/eBBkvbmi

🎉 Happy to share that our paper, "Examining the Threat Landscape: Foundation Models and Model Theft," has been accepted to the 34th British Machine Vision Conference (BMVC) 2024! 🎉 In our research, we explore the vulnerability of deep learning models fine-tuned from foundation models to model stealing attacks. While foundation models deliver high accuracy, they may not be the best choice from a privacy perspective when deployed in commercial APIs. Our findings reveal that these fine-tuned models are more susceptible to model stealing as compared to conventional CNNs like ResNets. This underscores a critical trade-off between privacy and accuracy in deploying models fine-tuned from foundation models in commercial settings. 🔐 A heartfelt thanks to Prof. Chetan Arora for his exceptional guidance and to my co-author, Deepankar Varma, for his invaluable contributions. 🙏 #BMVC2024 #ModelStealing #DeepLearning #AIResearch #Privacy #MachineLearning #FoundationModels