AIShield AI Security Offering from Bosch - Microsoft
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Our Stake Trust in AI: Safe, “Secure1”, Robust and Explainable We have to not only develop Success in creating AI would If Al is a black box, then AI, but build trust in AI as well be the biggest event in human people won't trust it. In a history. Unfortunately, it might connected world, however, also be the last, unless we trust will be essential. learn how to avoid the risks. Dr. Volkmar Denner Stephen Hawking Michael Bolle, Bosch CEO, on the ethics Theoretical Physicist Ex - CDO/CTO of Bosch of Artificial Intelligence 1 – our addition 2 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
Companies Deploying AI Cite Security Their Top Barriers to AI Implementation – AI Index Report 2021* Cybersecurity remains the only risk ,a majority of respondents say their organizations consider relevant.* Security for AI AI Security * Gartner (2019) and Mckinsey (2019,2020) found similar response from industry AI for Security 3 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
Adversarial Threats for AI/ML Models Evasion Poisoning Poisoning Spam content is embedded within Injecting malicious samples that an attached image to evade subsequently disrupt the retraining Training data analysis by anti-spam models process, e.g: Microsoft’s Tay chatbot Threat Actors: Cybercriminals, Motivation: Profit Threat Actors: Thrill-Seekers, Hacktivists Motivation: Satisfaction, Ideological Inference Inference Extraction Extraction Attempt to determine Probing ML system in order if the information of a to either reconstruct the certain record, e.g., of a person, model or extract the data has been part of the training data that it was trained on of a trained ML model or no Threat Actors: Insider Threat Actors: Cybercriminals, Threats, Cybercriminals Machine Learning Model Hacktivist Motivation: Discontent, Motivation: Profit, Ideological Profits Evasion 4 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
AIoT the next step in our I(o)T journey Security Threats across AIOT Cycle Bosch part of Consortium for Adversarial Threat Landscape for Artificial-Intelligence Systems Value Creation Connected Things and Services Data Artificial Intelligence Intentional Failures Unintentional Failures Perturbation, Reprogramming ML System, Model Natural Adversarial Exploit software Stealing, Recovering Training Data, Physical example, common dependencies Adversarial, Model inversion, Membership inference corruption, Value Stream Products/Services User Ideation Engineering Manufacturing Logistics AI Algorithms 1100 1010 0010 0101 1101 1010 1010 Data Flow Reward hacking, side Attacking ML Supply Distributional shifts Poisoning effects, Incomplete testing Chain, Backdoor ML AIoT is a revolution for value creation – better, faster, differentiating and secure 5 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
AIShield SaaS (Tool + API) Offering in AIOT Cycle Intrusion Detection and Prevention: Realtime protection, series or parallel configuration Preserve & Grow Deployment Operation Value Stream Products/Services User Ideation Engineering Manufacturing Logistics AI Algorithms 1100 1010 0010 0101 1101 1010 1010 Data Flow AIShield Data Model Training Data Processing AIShield Vulnerability Scanning: Analysis of Vulnerability Threat Intelligence Feed: New attacks identification End Point Protection: Defense Generation and integration Data Artificial Intelligence Platform agnostic AI Security Solution supporting various business models Value Creation Connected Things and Services 6 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
AIShield - Offerings Available Available Partially Available In roadmap Vulnerability Intrusion Threat Endpoint Protection Scanning detection prevention intelligence feed Targeted defense generation and Model theft vulnerability analysis for Real time prevention and Active threat hunting and incident Offering integration protecting against model various types of AI/ML models monitoring of new attacks report triggers extraction attacks ► Performs vulnerability assessment ► Generates targeted defense layer ► Protection against extraction and report generation supporting depending on type AI/ML of model, data attacks registered in the attack ► Report security incidents to SIEM >20 types of model, data type type variations (e.g.: image database via connectors Functional variations (e.g.: image classification, time series forecasting ► Ability to protect against new ► Threat hunting capabilities aided by Features classification, time series etc.) attack types and register Vulnerability analysis and active forecasting etc.) ► Able to integrate the generated defense telemetry data monitoring ► Able to ingest data, models from with original model for plug and play ► Frequent attack database ► Supports OSINT for AI Security various storage types operations in various configurations updates ► Support for various business Models (SaaS Tool, API SaaS for MLOps Integration) ► Available as APIs and Connectors to SIEM Usage Features ► Native support for automation ► Customization supported Pricing Subscription based enterprise license, Pay per use (call, instance/time based) Pay per incident detected/blocked, Pay per use (call, instance/time based) Customer AI Developers, PMs Security Experts, Operations, PMs Persona 7 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
AIShield Benefits – AI Product Teams Customers of Organization AI Product Team 1 4 ► Focus on building delightful ► Higher productive for users in cutting-edge products with organizations with simplified integration, security as added USP flexible and configurable workflows Organization AI Model Developer ► Enables customer 2 Benefits 5 ► Focus on building and retention, growth in newer improving models which are areas of AI security secure out of the box ► Securing Investment, IP ► Security engineering and Brand available with development workflow Organization Leadership Security Team ► Trust and Reliability of Products 3 6 ► New security paradigm with scale ► Strategic capabilities 8 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
Model Extraction Attacks What can you do in coming days? 1 Week (7 Days) 1 Month ( 30 Days) 1 Quarter (90 Days) 1 Year (365 Days) Start conversing with Assess the impact of Do a POC/Pilot with Create an organization- relevant teams on AI Model theft for your your read team and AI wide program under Security Topic AI/ML products team to validate MEA CISO to adopt new for your product security practices Create a list of AI/ML Make Organization Take a deep look at Engage vendors and product and assets leadership aware about MITRE Adversarial suppliers; also integrate which are offered to emerging threats and Threat Matrix and tools customers possible counter engage security measures researchers You can reach out to us AIShield.Contact@in.bosch.com 9 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
Engagement Process Overview of Process and Support Data BB – Attack GB – Attack Bosch POC Partner Bosch POC Partner Bosch POC Partner Request for Model & Model File and Blackbox attack vectors Extracted Model - Data for Greybox Provides Extra Data Validation Data Validation Data file (Gaussian, square block, Blackbox Accuracy provided Synthetic, combination) Validation Extracted Model - Blackbox GB- Attack Defense Benchmarking Bosch POC Partner Bosch POC Partner Bosch POC Partner Greybox attack vectors Extracted Model – Defense generation Integrated Defense Benchmarking Activities Sharing Results (Augmentation + Greybox attack (BB, GB) Model – Accuracy Blackbox attack vector) Accuracy Defense Integration Validation Extracted Model – Validation Greybox 10 Bosch | 2021-08-19 © Robert Bosch Engineering and Business Solutions Private Limited 2021. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.
THANK YOU Manojkumar Parmar Amit Phadke Program Director Product Manager Phone: +91 9513580024 Phone: +91 9148108633 Email: manojkumar.parmar@bosch.com Email: amit.phadke@bosch.com LinkedIn: manojkumarparmar LinkedIn: avphadke Twitter: @mparmar47 Twitter: @Maverics14
You can also read