Artificial Intelligence and privacy (Datatilsynet 2018)
A great report summarizing the main #ai features and #dataprotection challenges in the era of #artificialintelligence
It explains clearly
– Specialised vs Generalised #AI
– AI > ML > #deeplearning
– How ML works
– Classifications of #ML according to the way the model learn: supervised, unsupervised and reinforcement learning
– Problems associated with AI/ML: feature selection, breadth and depth of data processing
On transparency, it explains two models in an easily understandable manner: interpretable (decision trees) non-interpretable or black-box (neural network)
AI-GDPR issues
– algorithmic bias vs fairness principle (Art. 5(1)(a) GDPR)
– AI data repurposing vs. purpose limitation: art 5(1)(b)
– AI excessive data collection vs. data minimisation: art. 5(1)(c)
– use of non-explainable models vs. transparency: arts. 5(1)(a), 12-14, 22
Contrary to many publications, conflicts with the prohibition of non-ADM (art. 22) are clearly explained and illustrated with examples
Finally, it introduces some solutions and recommendations
Methods
– for reducing the need for training data (GAN)
– that protect #privacy w/o reducing the data basis (homom. encryption)
– for avoiding the black-box issue (explainable-AI)
0 comentarios