Charting The Future: White House Rolls Out a Landmark AI Executive Order

Limited memory machines are modeled after the way human neurons connect and share information in the brain. However, limited memory machines require large volumes of data to train the algorithms. With so much data and personal information on hand, businesses need to be able to ensure that the information will remain secure and protected.

In this respect, entities such as social networks may not even know they are under attack until it is too late, a situation echoing the 2016 U.S. presidential election misinformation campaigns. As a result, as is discussed in the policy response section, content-centric site operators must take proactive steps to protect against, audit for, and respond to these attacks. A second way to compromise data in order to execute a poisoning attack is to attack the dataset collection process, the process in which data is acquired.

ROI on AI deployment by empowering data driven decisions

For example, is the case of a user sending the same image to a content-filter one hundred times 1) a developer diligently running tests on a newly built piece of software, or 2) an attacker trying different attack patterns to find one that can be used to evade the system? System operators must invest in capabilities able to alert them to behavior that seems to be indicative of attack formulation rather than valid use. AI system operators must recognize the strategic need to secure assets that can be used to craft AI attacks, including datasets, algorithms, system details, and models, and take concrete steps to protect them. In many contexts, these assets are currently not treated as secure assets, but rather as “soft” assets lacking in protection.

For example, the social network may need to determine human involvement in and oversight of the system, such as by executing periodic manual audits of content to identify when its systems have been attacked, and then taking appropriate action such as increased human review of material policed by the compromised system. These suitability tests should be principled and balance potential harms with the need to foster innovation and the development of new technologies. The focus of assessments should include both current and near-future applications of AI.

Benefits of Domino

It is not certain that these problems could have been fully prevented through better planning and regulation. However, it is certain that it would have been easier to prevent them than it is to solve them now. The OECD’s Recommendation on Artificial Intelligence is the backbone of the activities at the Global Partnership on Artificial Intelligence (GPAI) and the OECD AI Policy Observatory. In May 2019, the United States joined together with likeminded democracies of the world in adopting the OECD Recommendation on Artificial Intelligence, the first set of intergovernmental principles for trustworthy AI. The principles promote inclusive growth, human-centered values, transparency, safety and security, and accountability.

Fines or penalties are meted to organizations for non-compliance due to a breach of data protection rules. Although current laws offer the basics for the protection of data privacy and security, they must progressively evolve to continue to remain relevant with the speed of technological advancements. The Framework is informed by IBM’s Financial Services Cloud Council which brings together CIOs, CTOs, CISOs and Compliance and Risk Officers to drive cloud adoption for mission-critical workloads in financial services. The Council has grown to more than 160 members from over 90 financial institutions including Comerica Bank, Westpac, BNP Paribas and CaixaBank who are all working together to inform the controls that are required to operate securely with bank-sensitive data in the cloud.

For example, CrowdStrike now offers a generative AI security analyst called Charlotte AI that uses high-fidelity security data in a tight human feedback loop to simplify and speed investigations, and react quickly to threats. Document management is also critical for education, state and local governments, and health care organizations. Customers like these that manage large amounts of structured and unstructured data and documents can consider deploying Quantiphi’s QDox, an intelligent document processing solution built by Quantiphi and powered by AWS.

(o)  The terms “foreign reseller” and “foreign reseller of United States Infrastructure as a Service Products” mean a foreign person who has established an Infrastructure as a Service Account to provide Infrastructure as a Service Products subsequently, in whole or in part, to a third party. (n)  The term “foreign person” has the meaning set forth in section 5(c) of Executive Order of January 19, 2021 (Taking Additional Steps To Address the National Emergency With Respect to Significant Malicious Cyber-Enabled Activities). (j)  The term “differential-privacy guarantee” means protections that allow information about a group to be shared while provably limiting the improper access, use, or disclosure of personal information about particular entities. In the end, AI reflects the principles of the people who build it, the people who use it, and the data upon which it is built. I firmly believe that the power of our ideals; the foundations of our society; and the creativity, diversity, and decency of our people are the reasons that America thrived in past eras of rapid change. We are more than capable of harnessing AI for justice, security, and opportunity for all.

In determining what attacks are most likely, stakeholders should look to existing threats and see how AI attacks can be used by adversaries to accomplish a similar goal. For example, for a social network that has seen itself mobilized to spread extremist content, it can be expected that input attacks aimed at deceiving its content filters are likely. Mitigation stage compliance requirements focus on ensuring stakeholders plan responses for when attacks inevitably occur. This includes creating specific response plans for likely attacks, and studying how the compromise of one AI system will affect other systems.

The first component of this education should focus on informing stakeholders about the existence of AI attacks. This users to make an informed risk/reward tradeoff regarding their level of AI adoption. Leaders from the boardroom to the situation room may similarly suffer from unrealistic expectations of the power of AI, thinking it has human intelligence-like capabilities beyond attack.

The growing use of AI technologies has pointed to the fact that governments around the world face similar challenges concerning the protection of citizens’ personal information. Partnering and sharing best practices better addresses these concerns in sustainable ways. Specified types of AI capabilities shall include generative AI and specialized computing infrastructure. In considering this guidance, the Attorney General shall consult with State, local, Tribal, and territorial law enforcement agencies, as appropriate. Through our long history of working closely with clients in highly regulated industries, we fundamentally know the challenges they face and built our cloud for regulated industries to enable organizations across financial services, government, healthcare and more to drive secured innovation.

(iv)   recommendations for the Department of Defense and the Department of Homeland Security to work together to enhance the use of appropriate authorities for the retention of certain noncitizens of vital importance to national security by the Department of Defense and the Department of Homeland Security. (C)  disseminates those recommendations, best practices, or other informal guidance to appropriate stakeholders, including healthcare providers. (B)  issuing guidance, or taking other action as appropriate, in response to any complaints or other reports of noncompliance with Federal nondiscrimination and privacy laws as they relate to AI. (ii)  any computing cluster that has a set of machines physically co-located in a single datacenter, transitively connected by data center networking of over 100 Gbit/s, and having a theoretical maximum computing capacity of 1020 integer or floating-point operations per second for training AI. Models meet this definition even if they are provided to end users with technical safeguards that attempt to prevent users from taking advantage of the relevant unsafe capabilities. (i)  The term “critical infrastructure” has the meaning set forth in section 1016(e) of the USA PATRIOT Act of 2001, 42 U.S.C. 5195c(e).

Our AI Principles, published in 2018, describe our commitment to developing technology responsibly and in a manner that is built for safety, enables accountability, and upholds high standards of scientific excellence. Responsible AI is our overarching approach that has several dimensions such as ‘Fairness’, ‘Interpretability’, ‘Security’, and ‘Privacy’ that guide all of Google’s AI product development. (xxix)    the heads of such other agencies, independent regulatory agencies, and executive offices as the Chair may from time to time designate or invite to participate. (ii)  Within 180 days of establishing the plan described in subsection (d)(i) of this section, the Secretary of Homeland Security shall submit a report to the President on priority actions to mitigate cross-border risks to critical United States infrastructure. (f)  To facilitate the hiring of data scientists, the Chief Data Officer Council shall develop a position-description library for data scientists (job series 1560) and a hiring guide to support agencies in hiring data scientists.

Read more about Secure and Compliant AI for Governments here.