The integration of Artificial Intelligence (AI) into cybersecurity has ushered in a new era of sophisticated threat detection, proactive vulnerability assessments, and automated incident response. As organizations increasingly rely on AI to bolster their defenses, the fundamental principle remains that the quality of the data on which they train these advanced systems directly links to their effectiveness. The old saying “garbage in, garbage out” (GIGO) holds true here; to unlock artificial intelligence potential – the power of pristine data is of the utmost importance.

Part 2 – The Perils of Poor Data Hygiene: Undermining AI Training and Performance
Neglecting data hygiene can have severe consequences for those who depend on AI for information. Data hygiene is directly correlated to the training and performance of AI models, particularly in the critical domain of cybersecurity. Several common data quality issues can significantly undermine the effectiveness of even the most sophisticated AI algorithms.
Missing Data
One prevalent issue is incomplete data sets, in particular the presence of missing values in datasets (https://www.geeksforgeeks.org/ml-handling-missing-values/). This is a common occurrence in real-world data collections due to various factors such as technical debt, software bugs, human errors, or privacy concerns. The absence of data points for certain variables can significantly harm the accuracy and reliability of AI models. The lack of complete information can also reduce the effective sample size available for training and tuning. This potentially leads to a decrease in a model’s ability to generalize. Furthermore and slightly more complicated, if the reasons behind missing data points are not random, the introduction of bias into some models becomes a real-world concern. In this scenario a model might learn skewed relationships based on the incomplete data set. Ultimately, mishandling missing values can lead to biased and unreliable results, significantly hindering the overall performance of AI models.
Incomplete data can prevent ML models from identifying crucial patterns or relationships that exist within the full dataset. Addressing missing values typically involves either:
- Removing data: deleting the rows or columns containing the missing elements. This comes with the risk of reducing a dataset and potentially introducing biased results if the reason for the data to be missing is not based on randomness.
- Imputation techniques: employing imputation techniques to fill in the missing values with guessed data. While this preserves the dataset size it can introduce its own form of bias if the guesses are inaccurate.
The fact that missing data can systematically skew a model’s learning process, leading to inaccurate and potentially biased outcomes, highlights the importance of understanding the nature of the missingness. The type of missingness are:
- Missing Completely At Random (MCAR)
- Missing At Random (MAR)
- Missing Not At Random (MNAR)
Understanding the reason at hand directly impacts the strategies for addressing this issue. Arbitrarily filling in missing values without understanding the underlying reasons can be more detrimental than beneficial.
Duplicate Data
Moving beyond missing data elements, another significant challenge is that of duplicate data within training datasets. While the collection of massive datasets has become easier, the presence of duplicate records can considerably impact quality and ultimately the performance and accuracy of AI models trained on this data. This can obviously lead to biased outcomes. Duplicate entries can negatively affect model evaluation by creating a biased evaluation. This occurs primarily when exact or near-duplicate data exists in both training and validation sets, leading to an overestimation of a model’s performance on unknown data. Conversely, if a model performs poorly on the duplicated data point, it can artificially deflate the overall performance metrics. Furthermore, duplicate data can lead to overfitting. This is where a model becomes overly specialized and fails to capture underlying patterns on new unseen data sets. This is particularly true with exact or near duplicates, which can reinforce patterns that may not be real when considering a broader data set.
The presence of duplicate data is also computationally expensive. It increases training costs with necessary computational overhead for preprocessing and training. Additionally, duplicate data can lead to biased feature importance, artificially skewing the importance assigned to certain features if they are consistently associated with duplicated instances. In essence, duplicate entries can distort the underlying distribution of a larger data set. This lowers the accuracy of probabilistic models. It is worth noting that the impact of duplicate data isn’t always negative and can be context-dependent. In some specific scenarios, especially with unstructured data, duplicates might indicate underlying issues with data processing pipelines (https://indicodata.ai/blog/should-we-remove-duplicates-ask-slater/). For Large Language Models (LLMs) the repetition of high-quality examples might appear as near-duplicates. This can sometimes aid in the registering of important patterns (https://dagshub.com/blog/mastering-duplicate-data-management-in-machine-learning-for-optimal-model-performance/). This nuanced view suggests that intimate knowledge of a given data set, and the goals of an AI model, are necessary when strategizing on how to handle duplicate data.
Inconsistent Data
Inconsistent data, or a data set characterized by errors, inaccuracies, or irrelevant information, poses a significant threat to the reliability of AI models. Even the most advanced and sophisticated models will yield unsatisfactory results if trained on data of poor quality. But, inconsistent data can lead to inaccurate predictions, resulting in flawed decision-making with contextually significant repercussions. For example, an AI model used for deciding if an email is dangerous might incorrectly assess risk, leading to business impacting results. Similarly, in security operations, a log analyzing AI system trained on erroneous data could incorrectly classify nefarious activity as benign.
Incomplete or skewed data can introduce bias if the training data does not adequately represent the diversity of the real-world population. This can perpetuate existing biases, affecting fairness and inclusivity. Dealing with inconsistent data often necessitates significant time and resources for data cleansing. This leads to operational inefficiencies and delays in project timelines. Inconsistent data can arise from various sources, including encoding issues, human error during processing, unhandled software exceptions, variations in how data is recorded across different systems, and a general lack of standardization. Addressing this issue requires establishing uniform data standards and robust data governance policies throughout an organization to ensure that data is collected, formatted, and stored consistently. The notion of GIGO accurately describes the direct relationship between the quality of input data and the reliability of the output produced by AI engines.
Here is a table summarizing some of what was covered in Part 2 of this series:
Data Quality Issue | Impact on Model Training | Potential Consequences |
Missing Values | Reduced sample size, introduced bias, analysis limitations | Biased and unreliable results, missed patterns |
Duplicate Data | Biased evaluation, overfitting, increased costs, biased feature importance | Inflated accuracy, poor generalization |
Inconsistent Data | Unreliable outputs, skewed predictions, operational inefficiencies, regulatory risks | Inaccurate decisions, biased models |
Part 3 will cover cybersecurity applications and how bad data impacts the ability to unlock artificial intelligence potential – the power of pristine data.