Machine learning algorithms are vulnerable to attack because of the way they learn statistical associations and can mimic human behavior. As a result, these models have no baseline knowledge, and they rely on data to learn. If data is tampered with or manipulated, an AI system can be made to behave like a Manchurian candidate. The following article outlines some of the main risks associated with artificial intelligence.
An adversary may capture physical equipment used by AI systems. Even the most advanced artificial intelligence systems based on deep neural networks are susceptible to attack. In addition to endogenous risks, AI systems require connectivity and flow of information. For instance, the use of digital farming systems relies on remote transmission and data transfer. Applications of AI for “smart cities” depend on real-time system actuation.
Changes in ecosystems may be accelerated by AI-related technologies. Because ecosystem services are vital to human development, AI-related innovations could negatively affect the welfare of social groups that rely on those ecosystems for survival. The potential negative impacts of increasing AI applications should be assessed carefully. There are a number of factors driving land-use changes. While most AI developments are not threatening to our daily lives, we should not underestimate their potential for social, economic, and environmental harm.
Increasing reliance on AI could threaten the right to privacy. Face recognition equipment, online tracking, and profiling can all use AI, and this can lead to unintended consequences. Further, AI-based algorithms combine various pieces of information, which could result in the creation of new information. A resulting data could be incongruous and potentially misleading. The use of artificial intelligence may threaten privacy and human interaction.
Poor data sets: A lack of quality data can result in biased AI-based results. For example, AI systems used in precision agriculture could use data from a small-scale farm, but may not be appropriate for that context. Incorrect management recommendations could result in lower yields for small-scale farmers. The rapid changes in ecosystems may also result in unintended consequences.
Security: Policymakers must understand the process of commercial AI development. Many companies will build proprietary AI systems, making it difficult for the government to pool resources and expertise. Additionally, because companies are building multiple systems at once, an attack on a single system will have limited impact. Moreover, AI systems can be hacked. Despite these challenges, policymakers must continue to monitor the progress of AI and consider how they will impact people and processes.
Lack of oversight: AI systems can be biased because of biases in data and design. Even in the best-case scenario, a bias in AI may be present. It may include factors such as race or ethnicity that do not appear in the data. This could lead to decisions that discriminate against certain groups. In addition, AI may influence the outcomes of criminal proceedings.
Unreliable AI systems: There are several major risks involved in artificial intelligence research. The main objective is to keep AI systems and algorithms safe and effective for society. Some AI systems may be hacked by adversaries and take over human tasks. These systems could even see vandalism as a green light. The future of artificial intelligence may depend on how well we manage these systems, as the security risks associated with AI are numerous.
Data Poisoning: Aside from data tampered with, AI systems can also be poisoned by an attacker. The attacker can manipulate the data or process used to train the AI system. The attacker can even “learn” a backdoor to control the AI system in the future. The worst case scenario for AI systems is when the attacker manipulates data.
Data Collection and Storage: An AI system’s collection process could be hacked by an adversary. The collection process of data is crucial to the success of the AI system. If it is not properly stored, the adversary could manipulate it and poison the AI system. This attack would cause the AI system to respond unpredictably and would make it useless for mission execution.
Input Attacks: Input attacks have many forms. Each attack has its own set of characteristics, but a taxonomy will help put an order into the vast array of possible input attacks. Input attacks can be classified by format – whether they involve real-world objects or digital assets. Input attack vectors are described in figure 2.
