Conf42 Quantum Computing 2024 - Online

Security Risks in GenAI from Revolutionizing Cybersecurity: The Role of Gen AI in Governance, Risk and Compliance

Video size:

Abstract

Discover how Generative AI is transforming cybersecurity in our talk on its integration into Governance, Risk Management, and Compliance (GRC). Learn how AI-driven automation enhances security efficiency by 40%, predicts threats 30% better, and reduces compliance violations by 25%. Join us.

Summary

  • Generative AI uses what it has learned from large amounts of data to create new, original content. In healthcare, generative AI can be used to analyze data from thousands of patients. In media, genai can be automatically used to generate new content for games or special effects. Ensuring security in AI systems means maintaining the integrity and fairness of what AI creates.
  • Data poisoning attacks basically target the training data of AI models. Attackers can skew an AI models output, making it unreliable or incorrect. The consequences of such attacks would be definitely severe. But what are the prevention and mitigation techniques for such an attack?
  • The next type of attack we are going to be looking at is called as model inversion attack. This is particularly concerning where models handle private data like personal identifiers or medical records. The techniques used to counter these attacks include limiting the amount of information the models output.
  • In risk management, Genai is used to predict and mitigate potential risks before they materialize. Application and example of GenAi in compliance include AI tools being automatically tracking and analyzing changes in regulations. Benefits of using genre in governments include enhancement of corporate efficiency and effectiveness.
  • Risk management in AI is a very crucial aspect for safe and effective use. Effective safe use of AI begins with identifying what can initially go wrong. Regular updates help protect against newly discovered vulnerabilities. Using AI security platforms to continuously monitor real time attacks is of paramount importance.
  • Ethical AI involves ensuring that these systems operate in a manner that reflects societal values, protects individual rights, and promotes fairness and justice. Some of the key concerns for ethical considerations include bias and fairness. Privacy is also an important key concern here.
  • As AI technologies become more integral to business operations, they also encounter a complex web of legal challenges. These involve compliance with existing laws and regulations and development of new laws to address emerging issues. Different countries may have different regulations regarding AI. AI systems operating across borders must navigate these kind of different legal landscapes.
  • With AI systems, if AI systems are being compliant with different regular frameworks, they are still being challenged. Challenges include regular audits and data management strategies. As technology keeps evolving, there's going to be the challenge of continuous education and awareness.
  • Security of generative AI is not a technical issue, but rather a broad concern that impacts public trust and ethical deployment of technology. One must be proactive, embedding security at every stage of AI development and deployment. This requires vigilance, continuous innovation, and the commitment to ethical practices.

Transcript

This transcript was autogenerated. To make changes, submit a PR.
We are gonna be basically going through an introduction to genaid and what it is, along with applications of Gen II across different facets of life. Then we will try to understand different key security risks in gen AI models. Then after that, we will go through real world case studies and how Gen AI is used in governance, risk management and compliance. We will then conclude by understanding the ethical and legal considerations of AI as a technology. So let's start by understanding what generative AI is. So, unlike traditional AI systems, which primarily analyze data to make decisions, like suggesting what movie you might enjoy next based on your viewing history, generative AI goes a step further. It uses what it has learned from different large amounts of data to create new, original content. This could be anything from a piece of music to a realistic image of a person who doesn't exist. So imagine a painter who learns by studying thousands of paintings, and then he or she starts creating unique art. That's similar to how generative AI works. It uses techniques based on advanced algorithms, specifically something called neural networks. So these neural networks are designed to mimic the way human brains operate. Another important technique involves what we call generative adversarial networks, or gans, where two neural networks contest with each other to improve the final output. Now, think of how this technology can be applied across different sectors. In healthcare, generative AI can be used to analyze data from thousands of patients to propose customized treatments, or simulate how drugs might work. Consider this scenario where there's sudden epidemic. So in the absence of lab rats, or maybe lab test rats that may be used, this software can be used to minimize the go to market time of a specific drug. This might be very helpful in saving thousands or millions of lives. Now let's consider the application of genai in media. In media, genai can be automatically used to generate new content for games or special effects in movies, thereby drastically reducing the time and cost involved in creative processes. Just think, the amount of time and people that are involved in creating an animation movie. So with Genai, these processes will be more efficient and more streamlined, and it will also help enhance and bring movies more closer to the fans who are expecting movies to be released at a faster scale. But why is security so crucial in Genai? So, as we depend more on AI to make decisions or create content, ensuring these systems are not only effective, but also secure and unbiased becomes paramount. If the training data is flawed, say it's biased against certain groups, those biases will definitely be reflected in the AI's output. Ensuring security in AI systems means maintaining the integrity and fairness of what AI creates, which is essential for gaining and keeping public trust. With that, let's go to different security attacks in Genai. First off, let's start with data poisoning attacks in Genai. Now, what is jet data poisoning attacked in Genai? So these kind of attacks basically target the training data of AI models, subtly inserting or manipulating information to affect the model's learning process, thereby causing it to make errors or bias judgments post deployment. So the mechanics for data poisoning attack would be by introducing corrupted, misleading, or specifically design data into a training dataset, attackers can skew an AI models output, making it unreliable or incorrect. For example, altering a street sign image used in training an autonomous vehicle's AI could cause the vehicle to misrecognize stop signs, as in signs. What would be the impacts and examples of data poisoning attacks? The consequences of such attacks would be definitely severe, especially in applications where like finance, where predictive models could determine credit worthiness, or in healthcare, where diagnostic accuracy is critical. We previously went through a traffic self learning traffic slash traffic example where self learning cars use the data to, to train themselves. The stakes here are much more higher because human life is involved. But what are the prevention and mitigation techniques for such an attack? So, solutions include robust data validation at the start before training them, making sure we use only trusted data sources to train the AI model, as well as implementing anomaly detection during the training phase to identify and mitigate suspicious data patterns. The next type of attack we are going to be looking at is called as model inversion attack. So what is a model inversion attack? So these kind of attacks basically focus on exploiting a model to relieve sensitive information about retraining data. So this is particularly concerning where models handle private data like personal identifiers or medical records. So the mechanics for such an attack. So by querying such an AI system with grafted inputs, attackers can reconstruct or infer data about individuals that were used in the model's training dataset, thereby effectively inverting the model's output to reveal the private information. By knowing certain private information, it's possible for the attacker to decipher the relationship and connections between different pieces of information. It doesn't take long to recognize most of our information is on social media. Just getting any one person identifying information can be dangerous and can be exposing the personal information of people. So what would be in the impact of such an attack, such as model inversion attack? So the impact of such an exam, such an attack would be in healthcare, for example, reconstructing patients faces from a model trained on medical images can be very dangerous because this could either misclassify or misreconstruct the patient's face, and thereby allowing the doctor to wrongly diagnose a patient for disease that they were not meant to imagine. If somebody is being diagnosed with another person's disease diagnosis, which could be life threatening for the specific person, and it has the ability to cause. So what would be the techniques for prevention and mitigation here? So, in this case, the techniques used to counter these attacks include limiting the amount of information the models output, as well as applying different privacy methods to obscure training data. The next type of attack we're going to be going through are called adversarial attacks in Genai. In adversarial attacks, they involve manipulating the input to the AI systems in a certain way that induce the model to make errors. So these attacks are often difficult to predict as they exploit the model's inherent weaknesses. So the mechanics for such an attack would be they are small, calculated changes to the input data, such that adding noise to image files that are impossible, imperceptible to the humans can lead to AI misclassifying these images or making erroneous decisions. An example here would be subtly ordering images that could trick a facial recognition system into misclassifying individuals, compromising security systems. Imagine what could happen if a facial recognition algorithm could classify an innocent person as a criminal and the criminal as an innocent person. This could have life changing consequences in both lives of both people. So what would be the prevention of such attacks? So, defenses for such attacks include adversarial training, where models are trained with adversarial examples to learn to withstand them, and implementing input validation to detect and filter out manipulated inputs. The last type of attack we're going to be going through is called backdoor attacks. So in backdoor attacks, basically what happens is when attackers introduce malicious functionality in a secret embedded manner during AI training phase, which later becomes active in response to a specific input trigger. The attacker basically introduces a trigger during the training phase that causes the model to act in a predefined malicious way when the model encounters the trigger during the operation. An example of a backdoor attack would be a surveillance system that could make it ignore specific individuals, effectively creating a security loophole. So the prevention and mitigation of such attacks that include rigorous inspection of training data, continuous monitoring of model behavior, and applying techniques to detect anomalies and models response. Now, let's go through different case studies where Gen A. So in case study one, we have a deepfake misuse. So the situation here is somebody is using deepfake technology to create realistically yet fabricated videos of public figures making contribution statements. In this case, it's a news anchor who belongs to a popular news channel that was created using AI. The impact of such deepfakes can spread, is that it can spread misinformation, it can influence public opinion and cause political instability. Along with this, it basically undermines the trust the public basically has in the media. So the mitigation of these attacks would be to carefully understand and analyze minute details in the videos. If in deepweight videos, its possible possible that the sound from the lips as well as the video is not always in sync. So these are very minute details one would not notice unless keenly observed. The development and application of deepwake detection tools is also helpful, and public awareness is basically crucial in understanding these types of attacks. In the second type of case study here is identity driven theft, where somebody is stealing someone's identity to perform a malicious task. So criminals use AI voice synthesis tools, in this case to impersonate a CEO, basically directing a subordinate to transfer funds to an unauthorized account. This leads to a big and huge financial loss, as well as reputational hit. This case study basically shows how criminals can use voice synthesizing a software to commit fraudulent activities, thereby exploiting the trust in human voice authentication. This case study shows the importance of different multiple factors of authentication that are needed. So some examples are maybe biometrics fingerprints, which are part of biometrics ubikeys, as well as two factor authentication techniques like phones, etcetera. So the need for setting up government governance frameworks for this kind of AI authentication system is particularly important in this case, so that it's easy to audit internal processes and prevent such type of fraud, and making sure that one or the organization itself is adherent to different regulatory standards. So the last case study here is AI manipulation. So in this case, an AI system was basically that used for trading was manipulated by poisoning the data which it used for training. So basically this caused the AI algorithm to make unpopular trades where the cyber criminals profited by making a hefty profit and the organization that was using the trading algorithm faced an irrecoverable loss. So the problem with these kind of attacks is it's because these are minor changes to the training data, it's very difficult to analyze and detect these algorithms. This shows the need of continuous auditing and detection that are needed along with anomaly detection systems in order to maintain the integrity and reliability of AI systems such as trading algorithms. So with that being said, let's go to application of GenaI in governance risk and compliance. So, in governance, generative AI is transforming how corporate governance is admins stirred by automating and optimizing decision making processes, thereby ensuring transparent and accountable governance practices. And a prime application example of this would be AI tools increasingly being used to analyze vast amounts of corporate data to identify different trends, forecast potential governance issues, and provide data driven insights to board members and executives. For instance, AI can automate the monitoring of compliance with corporate policies and regulatory requirements, alerting the management to potential non compliance and governance failures. The benefits of using genre in governments include enhancement of corporate efficiency and effectiveness. Basically, AI allows real time decision making with a higher degree of accuracy and less bias than traditional methods, provided the data that was used in training is itself fair. So it also supports a dynamic governance environment where strategic decisions are informed by up to the minute data analysis, enhancing responsiveness to market or internal changes. Let's go to the application of genai and risk management. So, in risk management, Genai is used to predict and mitigate potential risks before they materialize, using advanced analytical technologies to model risk scenarios and their potential impacts. The applications and examples of GenAi in risk management include being AI systems that are being adept at identifying different patterns and anomalies that may indicate risk, such as fraudulent activities or cybersecurity threats. For example, let's consider a financial services organization. A algorithms can predict credit risk by analyzing transaction patterns on customer behavior or more accurately than traditional models. The same can be said of banks. Banks use a algorithms, machine learning algorithms to basically identify the transactions we make and if that transaction passes above a certain threshold, they basically think that as fraudulent activity and many banks have the potential to block such activity. The benefits of using genai in risk management include allowing companies to anticipate and mitigate risks more effectively, thereby reducing the costs associated with losses and insurance. By enabling predictive risk management, organizations can allocate resources more efficiently and improve their overall risk posture. So lastly, let's look at GenaI in compliance. As for us visage of Gen AI in compliance, which is one of the key areas where AI can have a significant impact, especially regulated sectors like finance, healthcare and pharmaceuticals, etcetera. So AI systems here basically help ensure that compliance is maintained by monitoring regulations and automatically implementing changes. Application and example of GenAi in compliance include AI tools being automatically tracking and analyzing changes in regulations to help companies adjust their operations accordingly. In healthcare sector, for example, AI systems ensure patient data handling complies with the HIPAA or HIPAA regulations by automatically encrypting data and controlling access based on user roles, the benefits of use of AI in compliance. It basically reduces the likelihood of human error and the risk of non compliance, which could otherwise lead to hefty fines and legal issues. It also streamlines the documentation process, making audits more straightforward and less frequent. With that being said, let's go to risk management in AI. So as far as risk management in AI is concerned, AI is a very crucial aspect for safe and effective use. So effective safe use of AI begins with identifying what can initially go wrong, understanding the potential threats like data poisoning or adversarial attacks, as well as performing regular security assessments such as vulnerability assessments, threat modeling, helping anticipate and prepare for these kind of threats. So as long as security is included early in the face of development of a Gen AI model, it's safe to say that that model is likely to be more secure than models that try to incorporate security as an afterthought. Once the model has been trained, this likely has the potential to face different kind of attacks that we earlier saw. So employing tools like real time monitoring systems can alert to unusual AI behavior that might indicate a security issue. So encryption is a security technology here that provides data, that provides data by encrypting it with a strong algorithm, making sure only people who are able to access the specific data and interact with it will be able to interact it. So keeping AI systems and their components up to date with the latest security patches is also of paramount importance. Regular updates help protect against newly discovered vulnerabilities. Adversarial training basically teaches AI systems to recognize and understand disruptive inputs, enhancing their robustness. So utilizing AI security platforms to continuously monitor real time attacks is of paramount importance. Here we have different example security controls such as encryption access controls that we would be basically talking about and incident response plan is also of paramount importance, just because one needs to be ready to understand what would be the path one would take if a specific attack happens. So as we saw, incorporating adversarial examples during training is really helpful in helping the AI model distinguish between real pattern of data and then attack pattern of attack, basically where cybercriminals try to infiltrate the data. Let's go to ethical considerations for AI. So the rapid adoption of AI technologies brings to light numerous ethical considerations. Ethical AI basically involves ensuring that these systems operate in a manner that reflects societal values, protects individual rights, and promotes fairness and justice. Some of the key concerns for ethical considerations include bias and fairness. So AI systems can perpetuate or even amplify biases present in their training data, which lead to unfair outcomes that can discriminate against certain groups. For example, facial recognition technologies have shown lower accuracy rates for women and people of color, raising significant concerns about fairness and equality. The next point would be transparency and accountability. This is a growing demand for AI systems to be transparent about how decisions are made, what kind of data are being used to train them, especially during the training phase. So ensuring accountability when things go wrong is important, which requires clear guidelines on who is responsible. Is it going to be the developers, the users, or the AI itself? Privacy is also an important key concern here. AI systems often require vast amounts of data, which can include sensitive personal information. Ensuring that this data is handled security is of paramount importance so that privacy is maintained as a fundamental ethical requirement. Some of the strategies for addressing ethical concerns include implementing robust data handling and processing protocols to make sure there is fairness and reduce bias in the data that's used for training. The AI systems being developed they need to be developed with explainable AI features that allow users to understand and test the decisions made by AI. Basically, this is a way of saying the user should be able to understand how AI makes the decision so that it's not biased against certain people of color, women, or any kind of gender, etcetera. So regular ethical audits of AI systems are also helpful here as they help adhere to ethical standards throughout their life cycle. Let's go to the next section of the presentation, which is legal considerations for AI. So, as AI technologies become more integral to business operations and daily life, they also encounter a complex web of legal challenges. These involve compliance with existing laws and regulations and development of new laws to address emerging issues. Some example frameworks include a GDPR general data protection regulation, which basically imposes strict guidelines on data privacy and security, including AI systems that process personal data of EU citizens, European Union citizens. The other regulation we're going to be talking about is California Consumer Privacy act, which is similar to GDPR, but it provides consumers with specific rights regarding their personal data used by AI systems. This applies to the state of California only biometric Information Privacy act. This regulation basically regulates the collection and storage of biometric data, which is crucial for AI systems that use special recognition and other biometrics technologies. Some of the key challenges faced for legal consideration include intellectual property. Determining the ownership of AI generated content or invention pose significant challenges. Traditional laws here are not well suited to address whether an AI can be a copyright or a patent holder. Traditional laws were defined for publications. AI itself cannot be defined as a publication or a magazine or anything else because it's being generated by a software. And as technologies evolve, these kind of laws and regulations need to evolve to make sure they are being adapted to the evolving technology liability. As we, as we saw before, when AI systems cause harm, whether physically, financially or emotionally, determining the liability of a specific system can be complex. For instance, if an autonomous vehicle is involved in accidents, questions arise about whether the manufacturer, the software developer who provides the navigation system and the understanding, the self learning, understanding self driving system, or the vehicle owner itself should be held responsible for such an accident. The next part would be regulatory compliance. Different countries may have different regulations regarding AI, such as GDPR in Europe, as we saw, which imposes strict rules on data protection and privacy. AI systems operating across borders must navigate these kind of different legal landscapes. There is a new AI act that recently came that has been developed by EU. So as these regulations keep evolving to meet the legal evolving technology, it's a up to the organizations to make sure those that are operating across different borders must navigate and adhere to these legal landscapes. Some of the strategies for navigating legal challenges include engaging with legal experts to make sure that AI applications comply with relevant laws and regulations, participating in industry groups and standards organization to help shape the laws and regulations itself that are fair and practical, as well as implementing rigorous testing and compliance checks within the development process to mitigate potential legal issues before they arise. With that being said, let us look at some of the challenges currently facing AI systems. So with AI systems, if AI systems are being compliant with different regular frameworks, they are still being challenged. Challenges faced by organizations, basically that perform out of the border transactions. Some of these challenges include regular audits, which is either at a state level federal or a globe federal level, or a central level. Data management strategies. So as different nations have different regulations, data management becomes very important. Some regulations require data to be stored in state or within the country, or need certain regulations to be followed before they are being accessed outside the country. So, cross functional compliance teams, some of the challenges here include making sure working with cross functional compliance teams to make sure these organizations that operate across borders are adhering to different regulations frameworks, making sure that these organizations are evolving their systems to meet such regulations to keep in touch with what is needed for these technologies to stay safe and compliant. The last part is going to be continuous education and awareness. Of course, as technology keeps evolving, there's going to be the challenge of continuous education and awareness. So it's up to these organizations to devise a training plan, work with their different partners, and employees to make sure each and every one of them is up for the different evolution of technologies. With that being said, let's conclude this session by talking about what by summarizing what we saw and what we definitely how AI systems itself are beneficial to humans. So, to conclude, the security of generative AI is not a technical issue, but rather a broad concern that impacts public trust and ethical deployment of technology. One must be proactive, embedding security at every stage of AI development and deployment. So security needs to be deployed, considered as early as possible, rather than afterthought. The commitment to robust security measures must evolve as new threats emerge. This requires vigilance, continuous innovation, and the commitment to ethical practices. By doing so, one can ensure generative AI technologies are not only powerful and effective, but also safe and trustworthy. Let's take forward the understanding that security generative AI is integral to leveraging its full potential. We need to work collaboratively across industries and disciplines to uphold the highest standards of security and ethics, ensuring that it benefits the entire society without compromising the safety or integrity of the public. Thank you.
...

Akshay Sekar

Senior Technical Compliance Manager @ Intuit

Akshay Sekar's LinkedIn account



Awesome tech events for

Priority access to all content

Video hallway track

Community chat

Exclusive promotions and giveaways