Chapter 3. Perception of the Environment Author Martin Drasar Abstract This chapter discusses the intricacies of cybersecurity agents perception. It addresses

2025-04-30 0 0 291.09KB 14 页 10玖币

侵权投诉

Chapter 3. Perception of the Environment

Author: Martin Drasar

Abstract: This chapter discusses the intricacies of cybersecurity agents’ perception. It addresses

the complexity of perception and illuminates how perception is shaping and influencing the

decision-making process. It then explores the necessary considerations when crafting the world

representation and discusses the power and bandwidth constraints of perception and the underlying

issues of AICA’s trust in perception. On these foundations, it provides the reader with a guide to

developing perception models for AICA, discussing the trade-offs of each objective state

approximation. The guide is written in the context of the CYST cybersecurity simulation engine,

which aims to closely model cybersecurity interactions and can be used as a basis for developing

AICA. Because CYST is freely available, the reader is welcome to try implementing and

evaluating the proposed methods for themselves.

1 Background

Perception is a critical component of AICA and one of the few that cannot be omitted. Perception

provides information about the environment, communicates the results of the agent’s actions, and

shapes and influences the agent’s reasoning. While it may be possible to consider only the raw

data gathered from sensors as the perception, this narrow view does not appreciate the complexity

involved and only defers the issues of percept processing to other parts of AICA, such as the

decision-making engine.

Perception in AICA is as multifaceted concept as it is in biological systems. Even though the

artificial systems have the benefit of not being required to copy nature, many of the constraints

and drivers are universal. The raw percepts or stimuli go through a lot of preprocessing and

transformations before they can be subjected to the reason. Consider the optical illusion in

Figure 1. Our brain is hardwired to identify real-world objects, so we get thrown off because they

are not there. Moreover, it takes actual willpower to treat this image as just an image. The

perception mechanisms shape how we think about our environment, and the same goes for AICA.

Figure 1: An optical illusion.

There are multiple ways to conceptualize the perception in AICA. One possible way is in the

context of the DIKW pyramid, which conceptualizes the relation between data, information,

knowledge, and wisdom (Ackoff, 1989). This is depicted in Figure 2, where the perception

occupies the two lower tiers of the pyramid (data and information) but can sometimes venture up

to the knowledge tier due to its close relation with AICA’s world model. Another way we will

adopt in this chapter is a pipeline, as shown in Figure 3, consisting of four main parts: physical

sensors, logical sensors, transformers, and the world representation.

Figure 2: DIKW pyramid (Baldasarre, 2017)

Figure 3: A simple perception pipeline.

Physical sensors: are primarily out of the scope of AICA. Physical sensors process non-virtual

stimuli reaching the agent from the environment. Each of these sensors has specific operation

capabilities, requirements, and physical domain, but they all share the need for power. Therefore,

AICA using physical sensors must very carefully manage its power envelope.

Physical sensor examples: Temperature or pressure sensors and noise detectors could be

employed by AICA tasked with maintaining physical security inside a building. Perimeter sensors

could be used in outside deployment. Gyroscopes and lidars may be used within the context of

unmanned vehicles.

Logical sensors: in the context of this chapter, they are understood as a counterpart to the physical

ones. That is any source of data that rests within the software. A vast range of data can be fed to

AICA in this way. Ranging from its internal state measurements, host diagnostics, and network

measurements to open-source intelligence readings and even news feed. The only common

attribute of this data is that there is nothing in common. The data provided by logical sensors is

heterogeneous, with many dimensions, and can potentially require a large bandwidth to process.

These attributes go counter to the current reinforcement learning algorithms, so there is a need for

data reduction.

Logical sensor examples: Reading of running processes to gather information about the state of

AICA and the infrastructure it operates in. Network probe to gather information about traffic

within a guarded infrastructure. A periodic download of the CVE (MITRE) database to provide

updates to AICA’s knowledge base.

Transformers: provide means to reduce data complexity, dimensionality, and size. They ensure

the move from the data tier of the DKIW pyramid up to the information tier. They can provide

additional semantics to the data and serve as a heuristic that offloads a part of the logic that we do

not want the ML algorithms to discover. There are many different types of processors, arguably

more than types of data. The selection of transformers ultimately dictates how an agent perceives

the environment and how it can reason about it.

Transformer examples: Statistical aggregation and transformation of observed network traffic

(from packet traces to flows). Anomaly detection (from flows to events). Application of ML-driven

tools (from events to patterns).

World representation: is AICA’s representation of itself and of the environment it operates in.

A model of the world as it is being perceived. It is the foundation on which the agent chooses its

actions and against which their impact is evaluated. Currently, there exist no firm guidelines for

the design of state representation. If anything, it is considered an art by some because the

representation influences which algorithms can be used, how demanding the agent’s training will

be, and ultimately, what the agent can achieve.

Even though a pipeline is a fitting and easy-to-grasp concept, it gives an illusion of serial data

processing. However, the sensors are usually independent, and the same mostly holds for

transformers. As the data is being processed in parallel, delays, time skews, and interval

differences are bound to happen, as illustrated in Figure 4. The impact of these irregularities

strongly depends on the agent’s mode of operation and choice of algorithms. Passive observing

agents are largely unaffected because they can evaluate snapshots of the world state as the data

comes in. However, for active agents, this de-serialization can impact AICA’s efficiency by

providing only partial observations over a longer time, thus impacting both learning and acting.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Chapter3.PerceptionoftheEnvironmentAuthor:MartinDrasarAbstract:Thischapterdiscussestheintricaciesofcybersecurityagents’perception.Itaddressesthecomplexityofperceptionandilluminateshowperceptionisshapingandinfluencingthedecision-makingprocess.Itthenexploresthenecessaryconsiderationswhencraftingthewor...

展开>> 收起<<

Chapter 3. Perception of the Environment Author Martin Drasar Abstract This chapter discusses the intricacies of cybersecurity agents perception. It addresses.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Chapter 3. Perception of the Environment Author Martin Drasar Abstract This chapter discusses the intricacies of cybersecurity agents perception. It addresses

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: