12.4.8 : Cybersécurité
- Monday, (5:00 PM - 5:50 PM CET) Cybersecurity in the Age of AI: Navigating Threats and Innovations for VCs [SE63316]
- Sean Xiang : Co-Founder and CEO, Bloombase
- Ash Bhalgat : Senior Director of Cybersecurity Market Development, NVIDIA
- Denis Mandich : Co-Founder and CTO, Qrypt
- Ben Colman : Co-Founder and CEO, Reality Defender
- no record
- Monday, (5:30 PM - 5:55 PM CET) Accelerating NetworkX: The Future of Easy Graph Analytics [S61674]
- Mridul Seth : Core Developer, NetworkX
- Rick Ratzel : Senior Software Engineer, NVIDIA
- Slides
- NetworkX : open source project
- About Graphs
- Connectivity, shrtest path, route planning
- Everything can be a graph
- Almost 20 year old library
- Pure python dictionnary of dictionnary does not scale
- nx.betweeness_centrality() on 3.7 M Nodes, 16.5 M Edges, k=500, => 80 min on Intel Xeon Gold 6128 CPU, 45 GB RAM
- Backend : GOU nx-cugraph, CPU + OpenMP graphblas-algorithms, nx-parallel in pure python (in progress), scipy.sparse (not started but a good idea)
- Rapids.cugraph : no code change for GPU
- NETWORK_BACKEND_PRIORITY=cugraph ipython bc_demo.ipy
- nx-cugraph supports a lot a algorithms
- Better performancew with cache enable
- https://github.com/networkx/networkx
- https://github.com/rapidsai/cugraph
- cugraph scales to 1 B Edge, but nx-cugraph uses only one GPU for Now
- No link to pytock geometric, because networkx is an old school graph Analytics library an not a Maniche Learning graph library as PyTorch.geometric
- Tuesday, (4:00 PM - 4:50 PM CET) Build Secured AI Cloud Infrastructure at Scale [S62794]
- Tuesday, (5:00 PM - 5:50 PM CET) live Rethinking Cybersecurity in the Age of Generative AI: Emerging Generative AI Applications for Cybersecurity [S62696]
- 5B Internet user, several ZB generated
- Using GenAI to defend against GenAI
- Using GenAI to predict when GenAI is used for mallicious purposes
- You cannot have an hallucination which takes down a compagny
- Use of AI agent can reduce mean time to responce from 100 minutes to tens of seconds in some cases
- Even if 99% are false positive
- Generate documentation
- Takes 6 months to train a classifier, now it tool 2 or 3 days with Generative AI and Reenforcement AI
- Tuesday, (6:00 PM - 6:50 PM CET) How to Apply Generative AI to Improve Cybersecurity [S62173]
- Tuesday, (7:00 PM - 7:50 PM CET) Multi-Modal Integration of Deep Graph and Metric Learning for Optimal Malicious Behavior Detection [S61863]
- Abdul Rahman : Associate VP in the Artificial Intelligence (AI) Center of Excellence (CoE) in Advisory, Deloitte
- Adversarial model
- intention and attribution
- Deep graph and metric learning : increase the visibility of mallicious Behavior
- Single mode AI is a single tool
- Rule based tool => defined by human and need human to kept up to date
- PyG : Pytorch Graph Library
- Multi modal Behavior detection has very good result (qui ne veut rien dire)
- They also use Morpheus : Execution time per job goes from minutes to seconds (graph lo based method and learning method)
- Defining a deviation impiles a norm definition
- In 10000000 events there are 50 day threats, so traning data are unbalanced
- Representation of data can be done by a graph embeding
- Tuesday, (10:00 PM - 10:50 PM CET) GPU Power Play: Role of AI Democratization in Cyber Defense [S62820]
- Wednesday, (7:00 PM - 7:50 PM CET) The Intersection of AI and Security: What Cybersecurity Leaders Need to Know [S62433]
- David Reber : Chief Security Officer, NVIDIA
- Common strategies to help keep organisation safe
- More usage and more users => to protect
- Cybersecurity is a data problem
- 6x use of generative AI in Cybersecurity from 2022 to 2023
- Chat GPT : IPhone moment for AI
- Attacker : Right Once => Defender : Right Every Time
- Generative AI : Accelerate both Defender and Attacker
- Attack a digital twin of your network to see what will hapended
- What is vulnerable on this peace of worfware ?
- Is my network vulnerable to this new vulnerability ?
- Use AI on built fondamentals, not to replace everything
- How do I version PB of data ? (needed to train models)
- Still up in discussion
- Need standards
- How to respond to a malicious dataset ?
- Start with some AI use cases
- Granted access for only user with right permisions is not guaratee yet
- The only way to do that for now is to use several models trained on several part with diffrent critical data
- Learn the gaps before the date if too late
- The scale is huge (amount of data to be protected, to be learned, and the scale of attacks)