PowerSchool’s Post

Responsible, Ethical AI is the only way we can move forward with this powerful tool in education. The amazing benefits it provides are astounding, when done right. Check out our latest blogpost from Shivani Khanna Stumpf  outlining our stringent security and data governance of our AI tools. #AI #Education #K12

Who is really handling your data? As AI tools become more common in education, the convenience of having homework automatically graded can be appealing. But before you upload your students' assignments to an AI platform, it's important to ask: where is that data going? Many AI providers share data with third-party services to deliver these features, often without providing full transparency. Once student data leaves the original platform, it could be processed, stored, or even used to train future AI models — all without our knowledge. Students, especially young ones, are a vulnerable group, and their privacy deserves our highest attention. Before trusting AI tools, teachers and schools must demand clear answers about where student data goes and how it’s used. Privacy policies that don’t specify what happens beyond the platform raise red flags. In the rush to adopt AI, let's ensure we're not sacrificing the security of our students' personal information. If you want to make sure you know how to safeguard yourself, why not take my one-day training on ChatGPT? Find more information at https://www.aihorizons.nl Read my full article here: https://lnkd.in/eEjbE8wy #chatgpt #education #privacy #ai

This paper, titled “Machine Unlearning: A Comprehensive Survey” dated 13 May 2024, is an interesting review of the latest research of the emerging field of Machine Unlearning that aims to remove the contribution of specific data samples from already trained machine learning models. Context: The “right to be forgotten” is becoming relevant in the era of AI, to comply with regulations for data protection and privacy. Key Findings and Insights:   ·     Machine Learning contains a lot of randomness during the learning process and is difficult to access and to effectively remove the influence of the “to-be-erased” data.   ·     Centralized Unlearning (Exact Unlearning, Approximate Unlearning) and Distributed Unlearning (Federated Unlearning, Graph Unlearning) differs in challenges to unlearning. ·     Unlearning Verification is essential to assess the effectiveness of unlearning.   ·     Unlearning doesn't automatically guarantee privacy: it can paradoxically leak private information in some cases and introduce new privacy threats such as membership inference and data reconstruction attacks. ·     Unlearned models often perform worse than models retrained from scratch, a phenomenon called "catastrophic unlearning". You can find the paper here:

Are you interested in applying federated learning to real-world problems that involve security, privacy, and socio-technical aspects? Do you want to share your latest research work with an interdisciplinary audience of experts and practitioners? If yes, then consider submitting your paper(s) to the special session on Federated Learning Applications in the Real World – FLARW 2024, which will be held as part of the European Interdisciplinary Cybersecurity Conference (EICC 2024) in Xanthi, Greece, on June 5-6, 2024. The list of possible topics includes, but is not limited to: • Security or privacy-critical applications of federated learning in different domains (e.g., health, energy, finance, policing, e-government) • Human factors in federated learning applications (e.g., usability, attitude towards adoption) • Behavioural aspects of real-world federated learning systems • Legal aspects of real-world federated learning systems • New business models enabled by federated learning • Economic aspects of federated learning (e.g., incentivisation, economic modelling) The submission deadline is March 1, 2024. For more information visit the official link. https://lnkd.in/eq4m_k9y

The deadline of the special session has been extended to 15 March. Please continue to consider submitting to the special session. We particularly welcome #interdisciplinary #research papers. #federatedlearning #AI #machinelearning #security #privacy #cybersecurity #conference #Greece #EICC #interdisciplinaryresearch

The success of AI implementation hinges on robust infrastructure. Schools seeking to integrate AI must prioritize data quality and security.

The high cost of model training makes it increasingly desirable to developtechniques for unlearning. These techniques seek to remove the influence of atraining example without having to retrain the model from scratch. Intuitively,once a model has unlearned, an adversary that interacts with the model shouldno longer be able to tell whether the unlearned example was included in themodel's training set or not. In the privacy literature, this is known asmembership inference. In this work, we discuss adaptations of MembershipInference Attacks (MIAs) to the setting of unlearning (leading to their``U-MIA'' counterparts). We propose a categorization of existing U-MIAs into``population U-MIAs'', where the same attacker is instantiated for allexamples, and ``per-example U-MIAs'', where a dedicated attacker isinstantiated for each example. We show that the latter category, wherein theattacker tailors its membership prediction to each example under attack, issignificantly stronger. Indeed, our results show that the commonly used U-MIAsin the unlearning literature overestimate the privacy protection afforded byexisting unlearning techniques on both vision and language models. Ourinvestigation reveals a large variance in the vulnerability of differentexamples to per-example U-MIAs. In fact, several unlearning algorithms lead toa reduced vulnerability for some, but not all, examples that we wish tounlearn, at the expense of increasing it for other examples. Notably, we findthat the privacy protection for the remaining training examples may worsen as aconsequence of unlearning. We also discuss the fundamental difficulty ofequally protecting all examples using existing unlearning schemes, due to thedifferent rates at which examples are unlearned. We demonstrate that naiveattempts at tailoring unlearning stopping criteria to different examples failto alleviate these issues. #PrivacyProtection #UnlearningTechniques #MembershipInference #DataPrivacy #MachineLearning

It could be very helpful in addressing data privacy concerns and early adoption of LLMs.

We know that data is the key point to train ML and AI models. Best models derive from training sets that are the wider as possible and cover the necessary diversity of the topic we want to understand. We can see it in the LLM models: models trained on wider training sets have a greater capacity to learn complex patterns in the data and therefore gain better performances. The bad side is that higher needs of training data imply both an increase of the computational resources required and the need to search for new data that can contain sensitive informations. The federated learning framework helps us in handling these issues. From one side, federated learning can potentially decrease the need for computational resources compared to traditional centralized machine learning approaches. Indeed, in this framework the model is trained across multiple decentralized devices or servers, each holding a subset of the data, rather than on a single central server. This allows for the training process to be distributed, reducing the burden on any individual device and allowing for efficient use of available computing resources. Moreover, federated learning addresses privacy issues by enabling machine learning models to be trained on sensitive data without requiring the raw data to be transmitted or stored in a central location. This offers significant privacy benefits compared to centralized machine learning, but it should be combined with other privacy-enhancing technologies, like differential privacy or homeomorphic encription, to achieve strong end-to-end privacy protections. These two courses give a clear overview to better understand how this framework works in theory and practise.

The Federated Learning and Federated Computing will soon be getting massive adoption with courses of DeepLearning.AI and the highly esteemed Andrew Ng . Watch this space 👇