WEBVTT 00:01.200 --> 00:02.160 Hello everyone. 00:02.960 --> 00:14.160 In this particular section we will be learning about Microsoft Presidio for secure and responsible AI. 00:15.560 --> 00:24.360 You will learn why protecting personally identifiable information, also known as PII, is critical 00:24.360 --> 00:34.280 when deploying large language models and also understand how Presidio detects and anonymizes sensitive 00:34.280 --> 00:45.160 data and gain practical skills to secure your AI pipelines against data leakage and compliance violations. 00:45.880 --> 00:52.400 Having said that, let's first understand what is Microsoft Presidio? 00:53.920 --> 01:04.980 Microsoft Presidio is an open source framework designed to identify, assess, and anonymize personally 01:04.980 --> 01:09.460 identifiable information across various data types. 01:13.580 --> 01:24.780 Presidios acts as a privacy guardrail, and it sits between users and AI systems, ensuring that sensitive 01:24.780 --> 01:27.660 data never reaches LM APIs. 01:27.860 --> 01:30.700 Training pipelines or system logs. 01:32.100 --> 01:41.020 Some of its key features are that it has 50 plus PII recognizers for different entity types. 01:42.420 --> 01:53.460 It has NLP powered detection for customizable roles, flexible anonymization strategies, language agnostic 01:53.460 --> 01:58.620 architecture, and it's easy to integrate with the existing pipelines. 02:00.670 --> 02:03.990 So why is Presidio needed? 02:05.230 --> 02:12.270 Presidio helps in detecting protections. 02:16.270 --> 02:19.110 So why is Presidio needed? 02:21.790 --> 02:26.990 From the main challenges of PII is unstructured data. 02:27.710 --> 02:37.750 LM processes vast amount of free form text, making PII identification and protection exceptionally 02:37.750 --> 02:39.870 difficult without specialized tools. 02:42.110 --> 02:56.310 There are regulatory compliances like GDPR and CcpA, and there are severe penalties for non-compliance 02:56.310 --> 03:00.850 which can include hefty fines and legal actions. 03:02.250 --> 03:13.610 And then privacy by design is part of the architecture that should be maintained for modern AI applications, 03:14.250 --> 03:19.610 where privacy controls should be the heart of the application. 03:20.410 --> 03:28.010 Procedure enables proactive data protection rather than being reactive for each responses. 03:30.610 --> 03:32.410 So what does Presidio solve? 03:33.130 --> 03:43.530 Presidio directly addresses many of the pressing data challenges that emerges when integrating AI systems 03:44.050 --> 03:45.570 into enterprise workflows. 03:45.850 --> 03:54.370 The very first one is the PII detection in the free form text procedure excels at identifying PII in 03:54.370 --> 03:56.730 unstructured data from various sources. 03:58.130 --> 04:00.670 It also prevents data exposure. 04:01.550 --> 04:12.310 It ensures sensitive data is anonymized before it reaches downstream systems like third party API's, 04:12.430 --> 04:16.030 logs, or less secure environments. 04:17.990 --> 04:27.030 It standardizes the protection, provides a consistent framework for protecting PII across diverse data 04:27.070 --> 04:39.150 sets, and it secures AI workflows, which helps enables the safe use of Llms for sensitive tasks by 04:39.550 --> 04:46.110 ensuring that PII is never processed, stored, or revealed by the models. 04:49.310 --> 04:59.120 Presidio is directly connected to the OWASp for those who are deeply concerned with security, procedure 04:59.120 --> 05:07.760 offers reassurance by directly addressing critical vulnerability identified by leading industry standards, 05:08.200 --> 05:10.840 particularly the OWASp top ten. 05:12.960 --> 05:20.680 It aligns with the security standards, preventing sensitive data disclosures and implementing proper 05:20.680 --> 05:21.960 security controls. 05:24.680 --> 05:30.120 OWASp top ten focuses on sensitive information disclosure. 05:30.720 --> 05:38.640 One of the most critical risk in the AI application by sanitizing inputs and outputs procedure mitigates 05:38.640 --> 05:47.200 the risk of models revealing PII through completions or training data leakage. 05:48.920 --> 05:57.720 Now that we learned about Presidio and what it means for LM and AI systems, let's go ahead and understand 05:57.720 --> 05:58.640 its architecture.