What Is Edge AI? Navigating Artificial Intelligence at the Edge

Edge AI represents the deployment of artificial intelligence algorithms and models in an edge computing environment, which brings computational power and intelligence closer to where decisions are made, in part to offset a continuous communication stream between edge sites and the cloud. Edge AI enables devices at the periphery of the network to process data locally, allowing for real-time decision-making without relying on Internet connections or centralized cloud servers for processing, increasing computational speed, and improving data privacy and security. 

Understanding Edge AI

Edge AI is the convergence of multiple technologies, including artificial intelligence, Internet of Things (IoT), edge computing, and embedded systems, each playing a crucial role in enabling intelligent processing and decision-making at the edge of the network. Edge AI involves using embedded algorithms to monitor a remote system’s activity, as well as processing the data collected by devices such as sensors and other trackers of unstructured data, including temperature, language, faces, motion, images, proximity, and other analog inputs.

These remote systems can take many forms, including sensors, smartphones, IoT devices, drones, cameras, and even vehicles and smart appliances. The data collected from these systems serves as the input for edge AI algorithms, providing valuable information about the state of the system or its surroundings, allowing edge AI systems to respond quickly to changes or anomalies and understand the environment in which they operate. These edge AI applications would be impractical or even impossible to operate in a centralized cloud or enterprise data center environment due to issues related to cost, latency, bandwidth, security, and privacy.

Edge AI encompasses a wide range of use cases, including:

  • Autonomous vehicles. Edge AI enables vehicles to analyze sensor data in real-time to make split-second decisions for tasks such as object detection, lane tracking, and collision avoidance, without constant reliance on cloud connectivity.
  • Smart cities. Data from sensors and cameras deployed throughout an urban area can power various smart city applications, including traffic management, public safety monitoring, waste management, and energy optimization.
  • Agricultural monitoring. Edge AI supports precision agriculture by analyzing data from sensors, drones, and satellite imagery to monitor crop health, optimize irrigation, detect pest infestations, and perform real-time analysis of environmental conditions.
  • Industrial IoT. By deploying AI algorithms directly on manufacturing equipment and sensors, edge devices can monitor machinery health, detect defects, and optimize production processes without relying on centralized servers.
  • Healthcare monitoring. Edge AI supports remote patient monitoring and personalized healthcare by analyzing data from wearable devices, medical sensors, and imaging equipment to perform real-time analysis of medical data and alert healthcare providers to potential health issues.

Edge AI vs. Cloud AI

There are two primary paradigms for deploying AI algorithms and models: at the edge or in the cloud. Strategies to integrate systems that span cloud and edge sites are referred to as “cloud-in” or “edge-out”, with both having implications for performance, security, and operations.

Edge AI involves deploying AI on remote devices to enable real-time processing and decision-making at the network edge or in decentralized environments. These systems can largely analyze data locally, without relying on network connectivity or transmitting data to centralized servers, leading to lower latency and faster response times. Edge AI systems also keep sensitive data local, reducing the risk of privacy breaches or security risks associated with transmitting data to the cloud.

Examples of edge AI include autonomous vehicles that use locally deployed AI to analyze sensor data to make real-time driving decisions and smart home devices that use edge AI to process voice commands or monitor premises for intruders.

On the other hand, cloud AI is characterized by deploying AI algorithms and models on centralized cloud servers, allowing for large-scale data processing, training, and inference. Cloud resources bring significant computing capabilities, enabling complex AI tasks such as deep learning training or big data analytics that require massive computational power. Cloud AI solutions can easily scale to accommodate large volumes of data and users, making them suitable for applications with high throughput or resource-intensive requirements.

Recommendation engines such as those used by Amazon or Netflix to offer consumers new or alternative product choices based on extensive user data are examples of large-scale cloud AI systems that require substantial computational resources to function optimally.

Other AI use cases encompass both edge AI and cloud AI to meet specific customer needs. Real life examples include Sentient.io, a Singapore-based AI and data platform provider, which has developed the Sentient Marketplace, a hub of innovative AI services that allows businesses to easily integrate AI into their existing workflows. However, the marketplace’s rapid success presented several complex challenges, including the difficulty of operating and deploying AI services across distributed environments—on-premises, public cloud, private cloud, and at the edge.

When operating across multiple providers at customer sites, individual cloud-provider solutions may offer proprietary Kubernetes distributions, which can prove daunting for organizations that need to leverage these platforms in their respective cloud environments. Also cumbersome was the deployment process for Sentient’s AI models at customer sites, which called for setting up on-premises Kubernetes environments for each edge site, and handling updates and synchronization of new models manually. This resulted in increased operational complexity and inconsistent workflow orchestration and security policies.

Sentient.io partnered with F5 to offer turnkey, enterprise-grade AI “as a service” solutions to customers across a variety of verticals using F5 Distributed Cloud App Stack, an enterprise-ready Kubernetes platform that simplifies deployments across on-prem, cloud, and edge locations. The solution streamlined Sentient’s operations, reducing latency and enabling real-time AI processing at the edge. Delivering inference at the edge eliminates network and bandwidth constraints due to geographical location and ensures immediate delivery of inference to applications in real-time. This shift in model deployment enabled Sentient.io to deliver high performing AI applications to their customers with a faster time to value, optimize resource allocation, reduce overall operational costs, and natively integrate application and API security.

The collaboration also delivered significant cost savings over the previous process of managing multiple cloud platforms manually, which required dedicated teams and incurred substantial resource costs. With F5 Distributed Cloud Services, Sentient simplified operations, cutting costs by optimizing resources and simplifying application management, freeing up resources for other strategic initiativeo confirm.

Accessing Edge AI

Accessing edge AI involves deploying a combination of devices, technologies, infrastructure components, and integrations to enable efficient access and utilization of AI capabilities at the network edge. These include:

  • Edge devices. Embedded with sensors and microcontrollers, edge devices collect data from the physical world and can host edge AI models for local processing. Examples of IoT devices are smart thermostats, surveillance cameras, soil moisture monitors, and industrial sensors. Edge devices can also include smartphones and tablet computers, which not only sense their environment but can harness their processing power and connectivity to perform edge AI tasks.
  • Technologies. Operating AI systems at the network edge requires a number of specialized technologies, including trained algorithms and AI models that are optimized for deployment on resource-constrained devices. Edge AI frameworks and platforms are also available to provide tools and libraries to simplify system development and deployment. 
  • Infrastructure. Reliable network connectivity, whether wired or wireless, is required for edge AI devices to communicate with each other and to centralized servers when necessary, and can include such hardware components as edge servers, gateways, and routers. In addition, APIs are the linchpin in AI architectures, enabling different components and services to communicate with each other, and enabling them to exchange data and instructions.
  • Integrations. Edge AI systems must be able to integrate with existing networks and infrastructure to ensure data accessibility, enable scalability and compatibility with other components of the system, and ease management complexity.

Also, be aware of the following challenges and limitations to deploying and accessing edge AI.

  • Limited computing power and connectivity. Most edge devices have limited processing power, memory, and storage, which can restrict the complexity and size of AI models that can operate at the edge. Also, edge devices often operate in environments with limited options for network connectivity, which can also impact the responsiveness, performance, and reliability of edge AI systems.
  • Cost and availability. Many AI models benefit from workload accelerators such as Graphical Processing Units (GPUs) and Data Processing Units (DPUs) for faster processing, but GPUs, in particular, are costly and due to physical constraints can be too large for use in miniaturized form factors. This can limit the types of AI algorithms that can be deployed on edge devices and may require alternative optimization techniques.
  • Data privacy. Some edge AI systems generate and process sensitive or protected data locally, raising concerns about data privacy and compliance with regulations such as HIPAA or GDPR. Ensuring data privacy and compliance with legal requirements may require implementing appropriate data anonymization, encryption, and access control measures.
  • Device management. Deploying, monitoring, updating, and maintaining edge AI systems distributed across geographically dispersed locations can be challenging, and requires efficient management tools and platforms.


Edge AI Security Measures

Protecting data and mitigating security risks in edge AI deployments requires a holistic approach that emphasizes a multi-layered approach to security. While edge AI differs from traditional computing workloads in important ways, such as its ability to learn from data and evolve behavior based on experience, in terms of security requirements edge AI has much in common with more conventional IoT systems and shares many of the same risks, including: 

  • Malware and cyberattacks. Edge AI devices are susceptible to malware infections, cyberattacks, and remote exploitation if not properly secured. Implementing antivirus software, intrusion detection systems, and regular software updates should be part of every edge AI security strategy.
  • Network Security. Edge AI devices typically communicate with each other and with centralized servers over networks, making them potential targets for network-based attacks. Secure network communications with encryption, authentication, and access control mechanisms to protect data in transit and prevent unauthorized access to network resources.
  • Data Integrity. To maintain the accuracy and reliability of AI models and decision-making processes requires protecting the integrity of data processed by edge AI devices. Detecting and mitigating data tampering, manipulation, or corruption requires implementing data validation, checksums, and integrity checks to verify the authenticity and consistency of data inputs.
  • Physical Security. Edge AI devices are often deployed in remote or hostile environments, making them vulnerable to damage, physical tampering, theft, or vandalism. Physical safeguards, such as tamper-resistant enclosures or surveillance cameras, help secure devices from harm, manipulation, or unauthorized access.
  • API Security. AI ecosystems including plugins are facilitated through APIs, which are subject to vulnerabilities, abuse, misconfiguration, and attacks that bypass poor authentication and authorization controls.
  • Large Language Model (LLM) Security. LLMs and the relevant training and inference processes associated with decision making in generative AI-based applications are subject to numerous risks, including prompt injection, data poisoning, hallucinations, and bias.

For an in-depth examination of the security risks involved with deploying and managing AI systems based on LLMs, including edge AI applications, review the OWASP Top 10 for Large Language Model Applications, which promotes awareness of their vulnerabilities, suggests remediation strategies, and seeks to improve the security posture of LLM applications.  

Optimization Strategies for Edge AI

Because of its placement at the network edge or other remote locations, it’s important to optimize edge AI infrastructure for performance, resource utilization, security, and other considerations. However, optimizing for efficiency and performance for resource-constrained devices can be challenging as minimizing computational, memory, and energy requirements while maintaining acceptable performance often involves trade-offs.

Enhancing Performance in Edge AI

Several strategies exist to optimize computational performance at the edge while limiting energy consumption. Implementing power-saving techniques such as low-power modes, sleep states, or dynamic voltage and frequency scaling (DVFS) can help reduce energy consumption. Hardware accelerators like GPUs and DPUs can offload computation-intensive tasks from the CPU, improving inference speed. Use techniques such as dynamic batching, adaptive inference, or model sparsity to optimize resource utilization while maintaining performance. Less intensive tasks may be handled by CPU resources, underscoring the importance of resource pooling across highly distributed architectures.

Adapting Models for Edge Computing

Edge AI devices often have limited computational resources, making it necessary to deploy lightweight AI models optimized for edge devices. This can mean striking a balance between model complexity, accuracy, and inference speed when selecting the most suitable model for device resources and application requirements. Techniques such as model quantization, pruning, and knowledge distillation can help reduce the size of AI models without significant loss in performance.

Optimized Security at the Edge

The "dissolving perimeter" refers to how traditional network boundaries are becoming less defined due to factors such as mobile devices and cloud and edge computing. In the context of edge AI, the dissolving perimeter means that edge AI devices are usually deployed in remote and dynamic network environments at the network edge and operate outside of data center or cloud environments and beyond traditional perimeter-based security measures such as firewalls or intrusion detection systems. As a result, edge AI security has special requirements and must be optimized to protect against threats such as unauthorized access in isolated locations and across complex, distributed environments that make security management and visibility a challenge.

In addition, APIs provide the connective tissue that enables multiple parts of AI applications to exchange data and instructions. The protection of these API connections and the data that runs through them is a critical security challenge that companies must face as they deploy AI-enabled applications, necessitating the deployment of web app and API protection services that dynamically discover and automatically protect endpoints from a variety of risks.  

Security for Large Language Models

LMMs are artificial intelligence models based on vast amounts of textual data and trained to understand and generate natural language outputs with remarkable, human-like fluency and coherence. LLMs, which are at the heart of generative AI applications, are typically trained from input data and content systematically scraped from the Internet, including online books, posts, websites, and articles. However, this input data is subject to attack by bad actors who intentionally manipulate input data to mislead or compromise the performance of generative AI models, leading to vulnerabilities, biases, unreliable outputs, privacy breaches, and the execution of unauthorized code.

Among the top security risks to LLMs are:

  • Prompt injection. Attackers can manipulate LMM input prompts to influence the generated outputs and undermine the trustworthiness and reliability of LLM-generated outputs by generating biased, offensive, or inaccurate content.
  • Model poisoning. These attacks involve injecting malicious data during the training phase of LLMs to manipulate their behavior or compromise their performance. By introducing poisoned data samples into the training dataset, attackers can insert biases, vulnerabilities, or backdoors into the trained LLM model.
  • Model denial-of-service (DoS). These attacks target the availability and performance of LLMs by overwhelming them with malicious requests or inputs that can overrun request tokenization and LLM context window thresholds, causing slowdowns, disruptions, or service outages. These resource exhaustion attacks can lead to degraded performance or system instability, affecting the availability and reliability of the AI system, and compromising the ability of the model to learn and respond to user prompts.

To address these security challenges demands a multi-faceted approach that prevents prompt injections and employs techniques such as prompt sanitization, input validation, and prompt filtering to ensure that the model is not manipulated by maliciously crafted inputs. To counteract DoS attacks, create a layered defense strategy that includes rate limiting, anomaly detection, and behavioral analysis to detect and identify suspicious or malicious network activities. The industry is still evolving to effectively manage these risks, leading to rapid development of LLM proxies, firewalls, gateways, and secure middleware within application stacks.

The Future of Edge AI

Edge AI is part of a rapidly evolving set of technologies at the network edge, which is ushering in a new era of intelligent, responsive, and more efficient computing environments. These technologies, at the juncture of processor, networking, software, and security advancement, are unlocking new possibilities for innovation and transformation across industries. These edge computing use cases take advantage of real-time analytics and decision-making at the network edge, allowing organizations to process and analyze data closer to its source and improve response times for latency-sensitive applications or to ensure real-time delivery of content.

Distributing computing resources across the network edge also allows organizations to quickly adapt to changing workload demands and optimize resource utilization to improve overall system performance and efficiency. These possibilities are due in part to the evolution of purpose-built components for edge computing infrastructure, such as edge servers, edge computing platforms and libraries, and AI-on-chip processors that provide the necessary compute, storage, and networking resources to support edge AI applications.

Edge AI has played a pivotal role in driving the infrastructure renaissance at the network edge, and the integration of AI with the IoT continues to drive intelligent decision-making at the edge, propelling revolutionary applications in healthcare, industrial automation, robotics, smart infrastructure, and more.

TinyML is an approach to ML and AI that focuses in part on the creation of lightweight software ML models and algorithms, which are optimized for deployment on resource-constrained edge devices such as microcontrollers and edge AI devices. TinyML-based algorithms are designed to be energy-efficient, and capable of running inference tasks locally without relying on cloud resources.

In addition, compact and powerful processors such as DPUs, which are specialized hardware components designed to offload and accelerate data processing tasks from the CPU, are increasingly used in edge computing and AI/ML workloads, where the efficient processing of large amounts of data is crucial for performance and scalability. This efficiency is especially valuable in edge computing environments where power constraints may limit the use of energy-intensive GPU solutions.

Linking these innovations in an edge-to-cloud-to-data-center continuum is a new generation of networking solutions that enables seamless data processing, analysis, and observability across distributed architectures, including hybrid, multi-cloud, and edge computing resources. These networks will increasingly rely on APIs, which are essential components of edge computing platforms, as they facilitate communication, integration, and automation to enable seamless data exchange and synchronization within distributed computing environments. APIs also enable interoperability between diverse edge devices, systems, and services by delivering standardized interfaces, which also allows dynamic provisioning, management and control of edge resources and services.

In these wide-spanned distributed architectures, data can be securely processed and analyzed at multiple points along the continuum, ranging from edge devices located close to data sources to centralized—or dispersed—cloud servers located in data centers. This edge-to-everywhere continuum allows organizations to securely leverage the strengths of multiple computing environments and to integrate traditional and AI workloads to meet the diverse requirements of modern applications.

How F5 Can Help

F5 is the only solution provider that secures, delivers, and optimizes any app, any API, anywhere, across the continuum of distributed environments, including AI applications at the network edge. AI-based apps are the most modern of modern apps, and while there are specific considerations for systems that employ GenAI, such as LLM risks and distributed inference, these applications are also subject to latency, denial-of-service, software vulnerabilities, and abuse by bad actors using bots and malicious automation.

New AI-driven digital experiences are highly distributed, with a mix of data sources, models, and services that expand across on-premises, cloud, and edge environments, all connected by an expanding network of APIs that add significant security challenges. The protection of these API connections and the data that runs through them is the critical security challenge that companies must face as they deploy more AI-enabled services.

F5 Distributed Cloud Services offers the industry’s most comprehensive, AI-ready API security solution, with API code testing and telemetry analysis to help protect against sophisticated AI-powered threats, while making it easier to secure and manage multi-cloud and edge application environments. F5 Multi-Cloud Networking solutions offer SaaS-based networking with traffic optimization, and security services for public and private clouds and edge deployments through a single console, easing the management burden of cloud-dependent services and multiple third-party vendors. With F5 network solutions, you get accelerated AI deployments, end-to-end policy management, and observability for fully automatable and reliable infrastructure. 

In addition, the new F5 AI Data Fabric is a foundation for building innovative solutions that help customers make more informed decisions and take quicker actions. Telemetry from Distributed Cloud Services, BIG-IP, and NGINX combine to deliver unparalleled insights, produce real-time reports, automate actions, and power AI agents.

F5 is also releasing an AI assistant that will change the way customers interact with and manage F5 solutions using a natural language interface. Powered by the F5 AI Data Fabric, the AI assistant will generate data visualizations, identify anomalies, query and generate policy configurations, and apply remediation steps. It will also act as an embedded customer support manager, allowing customers to ask questions and receive recommendations based on model training of entire product knowledge bases.

By powering and protecting your AI-based apps, from the data center to the edge, F5 solutions provide powerful tools that deliver predictable performance and security so you can gain the greatest value from your AI investments.