The Good and Bad perspectives of Deepfakes

Deepfakes have emerged strongly only in the past few years. The name comes from Deep (Deep Learning) and Fakes (virtual, imaginary computer visuals). Deepfakes are an existing image or video created using artificial intelligence (AI), machine learning (ML) and deep learning techniques by the computer system that can mimic to look & feel 98% similar to our actual image and videos.

Deepfakes are created using AI and ML to create visual and audio content that can easily deceive people. The key technology and learning methods used come from deep learning and involve use of generative neural network architecture for training machine and generating output. It uses autoencoders, decoders and generative adversarial networks (GANs) from deep learning. 

Here is a brief overview of the Deepfakes are created and operated,

Deepfakes Creation

1.     Deepfakes work based on an autoencoder which is part of neural network in deep learning architecture. 

2.     The autoencoder reduces a person’s image to a low dimensional latent space. 

3.     The latent space captures the key features of person’s face and body postures etc. 

4.     This is then used by the decoder to reconstruct the image from its latent representations.

5.     The decoder works as part of generative adversarial network (GAN). The GAN trains the output generator consisting of the decoder and discriminator.

6.     The generator constructs the image from the latest representation while the discriminator checks and determines if the image constructed is matching to the source.

7.     The model then places the reconstructed image and features on top of the person’s video or image to mimic like the real person itself.

8.     As the entire architecture is functioning with AI, ML and Deep learning the model continues to take feedback, learn and evolve and it becomes so perfect that it’s almost impossible to differentiate between original source and deepfake reconstruct.

Now that we understand how deepfakes are created and operated, let’s understand why they are worse than good for society and organizations.

Deepfakes are not just be used on image and videos but also for mimicking audio. Deepfakes are not just be used for one person but even the entire group of people can be mimicked in an image, audio and visual. This is what makes them very dangerous as we can’t easily interpret fake Vs real on our electronic devices.

Deepfakes are generally used for mimicking others. They are often used by bad people targeting celebrities and leaders. The internet is full of bad examples of deepfakes. Some are even so shameful that we can’t even imagine.

There are some good sides of Deepfakes too. On the good end deepfakes can be used for,

1.     Reviving memories of people who have left the world and bringing them to life with virtual images and videos of them.

2.     Reviving memories of patients with short memory and memory loss.

3.     Marketing and Creative agencies are using deepfake to save time and costs on advertisements where the celebrity can shoot the ad just once and the audio content of the ad can then be mimicked and targeted towards respective target audience in their language making it more personalized and effective.

E.g., one such ad was from the leading food delivery company that is targeting its customers in different states and cities in different language using the Deepfake technology. The Deepfake celebrity is mimicked so clearly and accurately that the viewers can’t make out any difference and ad became very personalized and highly successful.

4.     Mobile Phone and website based apps for online learning of concepts by mimicking the actual creators of those concepts. 

E.g., Einstein teaching his theories to school and college students using online learning videos.

Deepfakes require strong internet connections and generally with weaker connection the lag can easily show up that the video is fake. But as the technology improves and countries move from 4G to 5G as well as the AI platforms and solutions becoming stronger along with Metaverse possibilities, it can lead to serious problems for many areas.

The solution to protect ourselves from bad sides of Deepfakes is to be careful,

1.     Using your photos, voice and videos on social media and storage platforms openly.

2.     We must ensure everything is well encrypted and securely stored. 

3.     Always be vigilant in day-to-day life and be alert and aware while using internet. 

4.     Stay away from using open (unsecured) and free access wireless network points. 

5.     Ensure you always use strong password on all your devices and change it periodically this will ensure your devices and accounts are always secured.

6.     Stay away from mobile apps and websites offering Deepfake free trial or even paid trials.

Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL) and many other new technologies have many benefits for organizations, society and people. But its very important to have a solid and secured good governance on how these new technologies can be used for the greater good of all.

Understanding Machine Learning and It’s How’s!

Machine Learning (ML) is part of artificial intelligence. In the background of most AIs there is machine learning at play. The term machine learning emerged in 1980s when computers started gaining their grounds. As the them name suggests machine learning, it is programs that help the machines learn and do statistical analysis and give results on probabilities and possible outcomes based on data.

Machine learning (ML) helps analyze large volumes of data based on several mathematical algorithms and share possible probabilities and possible outcomes within minutes. ML gives results based on algorithms and the results vary based on conditional variables and algorithm types we apply.

The results can’t be always called accurate because a lot varies based on what parameters and algorithms are applied to the data set. It also varies based on amount and quality of data used.

ML is used on large volumes of data which are impossible for humans to analyze with speed and accuracy. Machines can do this based on rules and programs set to produce results with speed. We can run the data through multiple algorithms and collect all the insights. Based on the insights collected further slice and dice can be performed and then we can decide the based way forward.

ML can also be used on any type of data, so not just text and numbers but also images and videos. In general, the ML process starts with,

1.     Cleaning the data set to ensure its consistency and no missing data. 

2.     Dividing the data set into training data set and data set for analysis.

3.     The training data set is used to train the machine learning algorithm on how to identify and what results to be expected. 

4.     The outcome of training data set is the trained model. To this model we feed the test data set to get the desired results. 

5.     The systems classify the test data set based on the patterns learned and logic used as part of the trained model.

6.     The prediction results can be verified with the trained model outcome to see the consistency. Based on the consistency the results and prediction outcomes can be accepted.

The machine learning is of 3 types. Namely Supervised Learning (SL), Unsupervised Learning (UL) and Reinforcement Learning (RL).

Supervised Learning (SL), as the name suggests, it is supervised means its trained with the labelled data set for identifying data sets and patterns to produce results based on the trained data set and other rules and algorithms to produce predictions. Supervised Learning (SL) works best for the classification problems, E.g., if we feed orange and other fruit images to the model as test data set along with the label specifying ‘its orange’ and for other fruits we ask it to label as “not orange”. Once the model is trained, we feed in the test data set with all fruits to check if it can identify and categorize fruits correctly in the outcomes.

Unsupervised Learning (UL), as the name suggests, it is unsupervised means it learns on its own from unlabeled data set. The algorithm automatically identifies various patterns in the data set and share the prediction outcomes based on the same. The full dataset if fed as input in the Unsupervised Learning (UL). UL identifies and creates its own group of data set based on various data attributes. E.g. if we feed multiple images of fruits to the data set, it can identify and group them based on attributes, like color, shape, weight, size etc. It will not know that this is a fruit or not but it will still efficiently group it and share the predictions and outcomes based on the data set attributes.

There is hybrid approach of combining Supervised and Unsupervised learning as Semi-Supervised learning where both SL and UL are used as certain attributes of dataset can be classified and trained while other remain unclassified and predicted using unsupervised learning. E.g. in the dataset if we know oranges clearly then we can train the model for identifying oranges accurately under supervised learning while all unknowns can be fed to unsupervised learning and that can group it and give predictions based on other attributes and features available in the data set.

Reinforcement Learning (RL), this works through a reward system where the system is rewarded for all right predictions and outcomes. This allows the system to learn based on its own actions that lead to rewards or penalties. Rewards are given for right predictions and penalties for wrong predictions. E.g., in a robotics factory every right action done by robot gets the reward and wrong actions gets the penalty. This allows the robotic system to learn the right and wrong actions based on the incentives (feedback) received.

Machine Learning (ML) has several algorithms and methods that are applied on the data set to build the complete model and share predictions that are impossible for human brain to manually achieve with in time limit of minutes and even hours.

Machine Learning (ML) is used for seemingly difficult predictions and probabilities to make effective decisions based on patterns identified by machine. Decisions like, 

1.     How many customers we must keep and which customer segment is most important for an organization in long term.

2.     Based on trends of past 2 decades when will be the predicted stock market crash.

3.     What age group purchases similar set of products

4.     We should tie up with which brand to combine our product bundles to increase sales

5.     Will Heart attacks increase and by what percentage

The results of the ML model can be verified using various performance metrics scores. In general, 6 performance metrics scores are checked to make a decision on viability of outcomes. The methods used are confusion matrix, F1 score, accuracy, precision, recall, and specificity.

As far as the tools are considered many of the ML capabilities are built in MS Excel. There are also standardized tools like Microsoft Azure ML, Google ML, IBM and many other preoperatory software tools. ML can be also managed using Python, R, Java, Julia, LISP programming languages.

Machine learning (ML) is heavily used in our day-to-day AI applications. ML is used by many market research, commercial, industrial organizations for identifying patterns and making difficult decisions. 

There is a lot of room for organisations for using ML to its fullest to make decisions. Opportunities are immense for all organizations and possibilities are wide spread across industries as we have big data available.

Understanding Artificial Intelligence and where it’s leading us!

It’s been almost a decade since we started recognizing the word artificial intelligence (AI) in our day-to-day life. AI has been around since many decades already but as it was more in the form known to us as computers and super computers only.

In the last decade it started to emerge more strongly as we started using artificial intelligence more widely in the form of search which after some years became voice enabled and even image enabled.

Artificial intelligence in simple terms is intelligence exhibited by machines and computers to do human like thinking, calculations, understanding and even day to day activities we carry out.

As the name suggests, artificial means not real human but it can more or less do everything that a human can do. Although we are still far from having 100% human like AI machines and robots but we are fast progressing in the direction of having it.

Artificial intelligence is of three types or we can call them having three different phases. The first one is Artificial Narrow Intelligence (ANI) followed by Artificial General Intelligence (AGI) and lastly Artificial Super Intelligence (ASI).

The ANIs (Artificial Narrow Intelligence) are initial use systems that perform normal tasks in one or two areas purely based on programmed rules. They can’t self-learn and can only execute specific tasks as per programmed. All their decisions and actions are pre-determined based on various options assigned to them. E.g., Google Translate, Email spam filter software, Chess app., Speech recognition etc.

The AGIs (Artificial General Intelligence) are more advanced than ANIs (Artificial Narrow Intelligence) and work seamlessly across multiple functional areas for problem solving, decision support and reasoning. E.g., Siri, Automatic Stock trading software that decide on multiple factors, statistics and news on when to buy and when to sell stocks for higher gains.

The ASIs (Artificial Super Intelligence) are not yet fully ready. As the name suggest super intelligence means human like abilities to multitask, decide, act, learn and relate things like humans do. There are trials ongoing and robots as well super computers are being made that can achieve this but we are still not full there and it might take still 1 to 2 decades to reach this level of AI solutions.

Artificial intelligence (AIs) solutions are very helpful in improving efficiency and effectiveness for almost all functions. Here are some examples,

1.     For industrial site solutions of robotics help improve production significantly. 

2.     For innovation and development, the solutions on product development, test automations and statistics decision-making help prepare viable solutions for meeting customer needs.

3.     For Commercial organizations, the solutions on robotics process automation, machine learning based statistical analysis for accurate decision making and real time dashboard with built automated decision guidance.

4.     For Consumers and Customers, the solutions on recommending right products based on their choices and know how in various areas and stages of life.

5.     For Customer Support, the solutions on full detailed understanding of their customer base and how to keep them as well as engage them to increase sales.

Similarly, there are also solutions based on industry specific needs. E.g., Your MD, Google DeepMind, Netflix, Snapchat, Apple Siri, Google Tools, Autonomous vehicles etc.

Artificial intelligence is used almost all of us in various shapes and forms. Simply put all our day-to-day apps we used in mobile phone and computers have AI built in to support us by enhancing our productivity and decision making.

AI has many positive sides but it also has worrying negative sides to it. The negatives are related to ethics and deception. How do we know that the AI is ethically deciding and making decisions as its all programmed, it can be designed to deceive. AIs can also be prone to hacking and manipulations of decisions and actions. There are still many open ends to AIs safe & secure usage and reliability. 

The other major worry for people is, if we automate everything with AIs and Robotics, then what will happen to current job market. There will be a significant shift of jobs needed as almost all of current manual jobs can be taken up by AIs with in this decade.

Possibilities are immense and when computers came there was a huge transition of jobs, a similar transition will take place in this decade with AIs becoming fully operational and even more when they evolve to Super Intelligence level in the coming decade.